Larsson Larsson Aberg Statistical Arbitrage Final Version 2002

A Market Neutral Statistical Arbitrage Trading Model
7 Development of Our Default Model

x 10
4.5
AFGX
AFGX 3m Moving Average
4 Our Model 3m Moving Average
Our Model
Net Position Short
3.5
2.5
Skr
1.5
0.5
−0.5
01/01/95 07/02/97 01/01/00 07/02/02
Date
Erik Larsson
Lars Larsson
Johan Åberg∗
13th March 2003
∗
Larsson, Larsson and Åberg are students at the Stockholm School of Economics and the Royal Institute of
Technology. The authors thank Patrik Säfvenblad (Head of Portfolio Management at RPM), Andrei Shleifer
(Professor of Economics at Harvard), and Johan Ahlberg (Datastream expert) for helpful comments. Email:
primus@kth.se (Erik), d98-lla@nada.kth.se (Lars) and 17774@student.hhs.se (Johan).
Abstract
The momentum effect is a systematic inefficiency in the market that can be exploited
by a trading strategy. This conclusion is supported by theoretical and empirical evidence.
But the academic research that tries to quantify the performance of this kind of strategy
often relies on a methodology that is too simplistic. The question arises what performance
a trader realistically could achieve in relation to the results presented in academic journals.
To answer this, we have written a computer program to run simulations with the added
realism of transaction costs and more advanced trading rules based on a wider array of data
than classic methodology allows. This has been done on Swedish stocks between 1995 and
2001. We then compare the simulation based on our own advanced model with a simulation
that emulates a simplistic methodology.
It is found that the negative impact on return of including transaction costs is outweighed
by the lower risk attributed to our more advanced trading rules, as indicated by e.g. Sharpe
and standard measures of risk. We can thus conclude that the momentum effect might be
even more attractive as a basis for a trading strategy than have been suggested in prior
academic research.
As an academic paper, we think that the methodology (our simulation platform) used to
obtain the conclusion in our thesis is more important than the conclusion itself. It is evident
that a good evaluation of any trading strategy requires more realistic simulations than is
commonplace in academia today.
Contents
1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Delimitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Outline and Reader’s Guide . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2 Previous Research 4
2.1 Theoretical Justification for Momentum Strategies . . . . . . . . . . . . . . . 4
2.1.1 Efficient Markets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1.2 Data mining and Spurious Data . . . . . . . . . . . . . . . . . . . . . 6
2.2 A Model of Over- and Underreaction . . . . . . . . . . . . . . . . . . . . . . . 7
2.2.1 Conclusions and Discussion . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Theoretical Motivation for the Stop-Loss Rule . . . . . . . . . . . . . . . . . . 14
2.4 Empirical Evidence Using CAR . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4.1 The Momentum Effect . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4.2 The Role of Volume, Fundamental and Sentiment Variables . . . . . . 17
2.4.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3 An Overview of Our Model and Its Context 19

3.1 Different Hedge Fund Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2 Our Model – an Optimization Problem . . . . . . . . . . . . . . . . . . . . . . 20
3.3 Technical Specification of the Computer Program . . . . . . . . . . . . . . . . 22
4 Data 24
4.1 Description of the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5 Methodology – Signal Generation and Risk Control 25

5.1 Signal Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.2 Risk Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.3 Assumptions in Our Simulations . . . . . . . . . . . . . . . . . . . . . . . . . 27
5.3.1 Basic Assumptions of Our Trading Model . . . . . . . . . . . . . . . . 27
6 Results - Portfolio Simulations 29

6.1 Results of Our Trading Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
6.1.1 Comments on the Results of Our Model . . . . . . . . . . . . . . . . . 32
6.1.2 Comparison with the Simplistic Momentum Strategy . . . . . . . . . . 33
7 Conclusions 35
A Appendices 36
A.1 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
A.1.1 Four Criteria of Robustness . . . . . . . . . . . . . . . . . . . . . . . . 36
A.1.2 Tables of Results for Robustness . . . . . . . . . . . . . . . . . . . . . 37
A.1.3 Discussion on the Robustness of Our Model . . . . . . . . . . . . . . . 43
A.2 Statistical Comparison of Our Model and the Simplistic Model . . . . . . . . 45
A.2.1 Volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
A.2.2 Return . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
A.3 Screenshot of the Program . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
A.4 A Day of Trading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
i
A.5 A List of the Swedish Stocks Used . . . . . . . . . . . . . . . . . . . . . . . . 49
A.6 Some Basic Financial and Statistical Concepts . . . . . . . . . . . . . . . . . 50
ii
List of Figures
1 Realizations of earnings over 600 periods . . . . . . . . . . . . . . . . . . . . . 10
2 The essence of the momentum effect. . . . . . . . . . . . . . . . . . . . . . . . 11
3 Histogram of our simulation of the overreaction effect . . . . . . . . . . . . . 13
4 Histogram of our simulation of the underreaction effect . . . . . . . . . . . . . 13
5 Conceptual view of the computer system. . . . . . . . . . . . . . . . . . . . . 23
6 A diagram over the development of our model, compared with the same in-
vestment in AFGX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
7 Histogram describing the distribution of the daily returns of our model . . . . 31
8 Histogram describing the distribution of the monthly returns of our model . . 31
9 The program in action . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
List of Tables
1 Illustration of how the investor changes her beliefs . . . . . . . . . . . . . . . 9
2 A brief summary of empirical research on the momentum effect. . . . . . . . . 16
3 The momentum effect in Germany, compared with Jegadeesh-Titman’s results
for USA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4 The momentum effect in Sweden 1980-1999 . . . . . . . . . . . . . . . . . . . 18
5 Hedge fund history at a glance . . . . . . . . . . . . . . . . . . . . . . . . . . 19
6 Categorization of hedge funds . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
7 Statistical Arbitrage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
8 Performance Jan. 1990 - Dec. 1999 of various hedge fund categories . . . . . 21
9 Our strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
10 Number of companies for which there exist various fundamental data at the
beginning and end of the test period . . . . . . . . . . . . . . . . . . . . . . . 24
11 The stocks used in the model . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
12 Our momentum-proxies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
13 Evaluation variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
14 Results of our trading model . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
15 Correlations of our trading model with different industries . . . . . . . . . . . 32
16 Results of the simplistic momentum strategy . . . . . . . . . . . . . . . . . . 34
17 Replication of earlier studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
18 Random rankings (median of 50 runs) . . . . . . . . . . . . . . . . . . . . . . 39
19 Transaction costs and risk control based on price variable . . . . . . . . . . . 40
20 Simulations from 1995 in Sweden . . . . . . . . . . . . . . . . . . . . . . . . . 41
21 Changes in our default model . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
22 Hypotheses testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
23 Swedish stocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
iii
1 INTRODUCTION 1
1 Introduction
1.1 Background
Technical analysts have always been the neglected outcasts of the world of academic finance.
Business students learn in their first year that any attempts to study historical data in search
of that elusive gold mine are futile. The market quickly abolishes such pretensions. And it is
not hard to see why the lecturer has that sarcastic tone in his voice while preaching the laws
of finance and their relentless effect on the pity chartist. Anyone who has read a book from
the 80’s on charting techniques knows why. Pen and paper are the tools, and with them one
will find specific patterns of stock prices emerging and then it is just to fill in the blanks and
the order book. No statistical or theoretical foundation, no portfolio view or risk control.
But often with a closing chapter on the link between the stock market and astrology.
This tarnished image did not disappear even as the upholding arguments of the efficient
market hypothesis (hereafter EMH) came under question in the 80’s; the assumptions of the
rational investor and the effect of arbitrage on bringing back prices to their fundamental
value. A tidal wave of theory and empirical evidence has since then given the chartists
right on their main point: historical data do provide some information valuable in predicting
future prices. The reason being irrational investors and limited arbitrage.
The old school of technical analysis has not changed that much though except for better
data and computing power, but the professional community of investment management did
change. In conjunction with the surge of alternative investments in the 90’s, many popular
strategies were based on the new paradigm, trying to exploit systematic inefficiencies in the
market. And the question of where the line between arbitrage profits and risk premium goes
may be one of semantic rather than practical importance, especially in the more number-
crunching inclined strategies.
In any case, the explosive growth of the hedge-fund industry, the availability of good
data and more powerful computers, have moved the frontiers for the statistically interested
investor and manager. The situation today is indeed horizons away from the pen and paper
chartist of the 80’s. Our own skills in programming and mathematics intersect with this new
development and the gold-shimmering prospect of finding an untouched deposit of a steady
stream of profits. This is of course the ancient romantic dream that have left many fortune
hunters in dismay and destitution. So a more realistic goal for this master thesis becomes
an attempt to use some of the new theory and empirical research that has come under the
name of behavioral finance in recent years to construct a trading model which performance
will be tested.
Specifically, we are looking at a strategy categorized as market neutral statistical arbit-
rage. It is an equity market neutral strategy of interest for investors who want to have an
asset with low market correlation. This strategy has caught the attention of many aca-
demicians lately, as being a good example of a systematic inefficiency on the market. The
methodology used when analyzing this phenomenon in most academic papers, called Cumu-
lative Abnormal Return (CAR), is straightforward and easy to use. However, it is also very
rigid and simplistic.
To assess the true potential of exploiting the momentum effect, we argue, one must simu-
late trading in a way that is much more realistic and that combines a variety of trading rules.
We have done this by constructing a simulation platform, a computer program, amounting
to a total of 220 pages of Java code.
1 INTRODUCTION 2
1.2 Purpose
The purpose of this Master Thesis is to assess the merits in terms of risk and return of a self-
made trading model based on the momentum effect. This is however not done in an absolute
way, instead we want to see how performance changes in our sample when we introduce both
transaction costs and more advanced trading rules in relation to the naive (or simplistic,
see Appendix A.6 for a definition) strategy common in academic papers on the momentum
effect. We can then determine whether it is feasible to exploit the momentum effect in a
more realistic setting.
As a second, but in our view more important objective of this thesis, we want to demon-
strate the simulation platform we have developed, and advocate its use as a superior meth-
odology when evaluating trading strategies.
1.3 Delimitation
We limit the scale and scope of our thesis in the following areas:
• Our data is limited to 6 years of Swedish financial and fundamental data (see further
Section 4). We have however stock quotes for more than ten years in Sweden, Germany,
France and the United Kingdom. This data will be analyzed in the Appendix to
estimate the robustness of our model and the underlying market inefficiency it tries to
exploit.
• The technical details of our model and simulation platform are not the focus of this
thesis and will not be examined in depth. It is not possible with the information in this
thesis to replicate the simulation platform or the exact algorithms that constitute our
own trading model. This would have required a lot of technical documentation that we
have chosen to withhold.
• We have still not optimized our market neutral statistical arbitrage model in a way
that would be recommendable in a real life setting. For example; we are fully invested
at all times, we do not consider an optimal hedge, and we do always include 10 long
and 10 short positions in our portfolio.
Further clarification and justification for the focus of the model within these limitations
are given in more detail in Sections 3 and 5.
1.4 Outline and Reader’s Guide

To our knowledge, no other master thesis in Sweden has been written in an attempt to
produce a full-fledged trading system. This should be seen in light of the fact that the
required computing power and data quality have not been available more than a few years
for student use.
What we gain in realism is lost in simplicity, and thus to some extent also in replicability,
transparency and focus. But realism is an objective of this thesis. As is our critique of
simplistic methodologies, such as CAR, and their equally simplistic conclusions that do not
extend beyond a mere theoretical domain, often without bearing on the real world.
Our thesis is structured as follows:
• Previous Research. First of all, we want to establish a theoretical and empirical
foundation for the inefficiencies (in an EMH sense) we want to exploit. We provide a
theoretical model that tries to explain the phenomenon of momentum in stock prices.
Also, we highlight some empirical evidence on the momentum effect from different stock
1 INTRODUCTION 3
markets, using the classic CAR methodology. Furthermore, we look at other variables
that could be used in the model, such as sentiment and fundamental variables.
• An Overview of the Model and its Context. Next, we look at our model in
the context of different trading strategies used by hedge funds and make some notes
on the technical structure of our simulation platform. This section does not lie on
the otherwise straight line between purpose and conclusions, but is important for the
general understanding of this thesis.
• Data. The data used as input in the model is described in this chapter.
• Methodology – Signal Generation and Risk Control. This chapter describes how
we turn the input into output, i.e. a market neutral portfolio. Signal generation deals
with how buy and sell signals are generated from the data. Risk control is necessary in
order to optimize the risk-return profile on a portfolio level, by taking into consideration
the risk implications of the generated buy and sell signals.
The word methodology here is referring to how we derive the algorithms used when
deciding what, when and how many stocks should be bought or sold at a specific mo-
ment. The methodology in a wider sense is referring to the Java program, or simulation
platform, that we use to implement the algorithms.
• Results – Portfolio Simulations. The model’s output is statistically analyzed and
then compared with a simplistic strategy with static rebalancing and only price as an
input variable.
Finally, we conclude on our findings. In the Appendix, we have an extensive discussion on
the robustness of our model, based on out-of-sample tests of our model using price data from
Germany, France, UK and Sweden. Basic financial and statistical concepts used in this thesis
are explained in Appendix A.6.
2 PREVIOUS RESEARCH 4
2 Previous Research
This section begins with a theoretical background to our experiments, and continues with
an illustrative model of irrational behavior. The section ends with a digression on our use of
the stop-loss rule and a brief review of previous empirical studies.
2.1 Theoretical Justification for Momentum Strategies

The ambition of this section is to give a theoretical motivation for the momentum effect.1
This motivation will be supported by Section 2.4, where we also deal with aspects of the
momentum effect that are of more practical use to our trading model. This theoretical
section is included since the momentum effect is of utmost importance to our trading model.
We will thus first present the main lines of critique of momentum, and then the answers to
these issues in turn.
Those who oppose the existence of the momentum effect argue mainly along one or both
of the following two lines:
The markets are efficient According to the EMH, a continuation of stock prices not based
on changes in fundamental values is not possible, or at the most a temporary anomaly.
Data mining Another explanation is that the effects discovered empirically simply are the
results of data mining, i.e. the data is spurious.
2.1.1 Efficient Markets

The EMH has a long history in finance, and its proponents are some of the most prominent
figures in financial economics such as Eugene Fama at University of Chicago.2 To start
off this section, we first discuss the concept of efficient markets and the implications they
have for momentum strategies, since our trading system would not work if the markets were
efficient, even if only in the weak form (see below).
In words, the EMH states that the current market price reflects the assimilation of all
the information available. This means that given the information, no prediction of future
changes in the price can be made. As new information enters the system, the unbalanced
state is immediately discovered and quickly eliminated by a ”correct” change in the market
price, i.e. the momentum effect should not be possible.
In Fama (1970), the concept of efficient markets was for the first time formalized, even
though the mathematical foundation had been put forth earlier in Mandelbrot (1966). Fama
expresses the non-predictable characteristic of market prices formally as
E(p̃j,t+1 |Φt ) = [1 + E(r̃j,t+1 |Φt )]pj,t , (1)
where pj,t is the price of security j at time t, rj,t+1 is the one-period percentage return
(pj,t+1 − pt,j )/pt,j , Φt is the information reflected at time t, and the tildes indicate random
variables. The expression E(A|B) indicates the expected value of A given that event B has
occured. Fama then goes on to state explicitly that this rules out any possibility of trading
systems based solely on information in Φt that have expected profits or returns in excess of
equilibrium expected profits or returns.
1
The momentum effect is here defined as ”the persistence in the returns of stocks over horizons between three
months and one year” (Grinblatt and Han, 2002, p. 1). The typical set-up of a momentum strategy is to go long
in past winners and take a short position in past losers.
2
We will sometimes refer to the champions of the EMH as ”traditionalists”.
According to Fama, the EMH has three different shapes, based on how ”large” the in-
formation set Φt is; the strong form (where groups of investors have more information than
others, e.g. insiders), the semi-strong form (where the information set includes all the obvi-
ously publicly available data such as volume data, profit prognoses and sales forecasts), and
the weak form (where the information subset is just historical price or return sequences).
He concludes that the weak and semi-strong forms are heavily supported by empirical tests,
while the status of the strong form is more ambigious, since it is very hard to test statistically
due to shortage of data. He also discusses the existence of autocorrelation in price sequences,
which is of utmost importance for momentum strategies, but concludes that they are either
insignificant or too small to make any economic sense. Furthermore, he mentions trading
strategies that beat buy-and-hold on a consistent basis when not regarding transaction costs.
It is worth noting that we have tested our model with transaction costs.
The essence of Fama’s paper is that one cannot systematically beat the market, using e.g.
a momentum strategy. So why would a momentum strategy work? The theory presented
above is theoretically sound, the problem is that the assumptions upon which the theory rests
are not plausible nor applicable in reality. The most fundamental of these assumptions is
that investors behave rationally, or, if they do not, that the deviations of irrational investors
cancel each other out, which is the view expressed in e.g. Fama (1998). Another ”efficient
markets”-explanation is that the irrationality is exploited by arbitrageurs, something which
has been put forth by de Long et al. (1990).
But research done the last 30 years has shown that investors in fact are not strictly
rational nor possess unlimited computational capacity (Shiller, 1998). Instead of trying to
compute expected values and arrive at sensible strategies through the use of Bayes’ rule,
investors commonly rely on various heuristics.3 Particularly interesting for our thesis is the
representativeness heuristic, by which is meant the ”tendency for people to try to categorize
events as typical or representative of a well-known class, and then, in making probability
estimates, to overstress the importance of such a categorization, disregarding evidence about
the underlying probabilities” (Shiller, 1998, p. 22). A manifestation of this somewhat ab-
stract idea is the tendency of people to see patterns in strictly random sequences of numbers.
People are also victims of various biases, of which conservatism bias is one of the most im-
portant. People with a bias towards conservatism tend to underweight new information in
updating their weighting of new information (Jegadeesh and Titman, 2001). For a thorough
exploration of these issues, we refer the interested reader to e.g. Shiller (1998).
The above biases and rules of thumb all point to deficiencies of the EMH, but what
about the proposition that the effects of irrational investors trading according to various
heuristics or biases cancel each other out? Experiments done by Kahneman and Tversky
have shown that individuals tend to deviate systematically, and deviations will thus have a
real impact (Shleifer, 2000).
What specifically do the heuristics and biases lead to? First and foremost, they generate
over- and underreaction to e.g. news. These reaction effects are essential to the momentum
effect, which will be shown below. We describe these two effects formally as follows:
Overreaction:
E[rt+1 |zt = G, zt−1 = G, . . . , zt−j = G] < E[rt+1 |zt = B, zt−1 = B, . . . , zt−j = B] (2)
3
Bayes’ rule states that, if Ai , i = 1, . . . , n and B are events, then
P (Ai )P (B|Ai )
P (Ai |B) = Pn ,
j=1 P (Aj )P (B|Aj )
see e.g. Råde (1999). In words, investors following Bayes’ rule will weight information after how precise it is.
Underreaction:
E[rt+1 |zt = G] > E[rt+1 |zt = B], (3)
where zt is the news that the investor hears in period t, G is good news, B bad news and
rt is the period t return. Below, we will give examples of good and bad news. In words:
overreaction occurs when the average return (an approximation of expected return) following
a series of announcements of good news is lower than the average return following a series
of bad news announcements. Underreaction in its turn occurs when the average return on
the company’s stock in the period following an announcement of good news is higher than
the average return in the period following bad news (the notation as well as the definitions
are from Shleifer (2000)).
The momentum effect and the phenomena described above that cause it, were first doc-
umented in Bondt and Thaler (1985), where empirical support for the overreaction effect is
found.4 Evidence of the underreaction effect can be found in Jegadeesh and Titman (1993).
The traditionalists have also tried to find other explanations than the irrational behavior
of investors to the momentum effect. Fama and French (1993, 1996) have e.g. tried to relate
the phenomenon to risk premia on investing in value stocks and small stocks, which would
imply that the EMH is not violated. They take the comovements of stock prices in these
categories as evidence that the securities share some common fundamental risk. But this has
not been shown directly (Shleifer, 2000), and instead there could be a behavioral explanation,
as shown by e.g. R. La Porta and Vishny (1997). A case can be made that similar stocks are
influenced by the same investor sentiment, so that the stock prices are correlated. The upshot
of this is that the comovement of stocks does not have to be explained by risk premia. Another
setback for the EMH is that an implication of the rational explanation of the value/glamour
evidence is that investors should expect lower returns when investing in glamour stocks
because of those stocks’ risk characteristics. But evidence points the other way, i.e. that
investors are too optimistic about the future returns of growth stocks (J. Lakonishok and
Vishny, 1994).
2.1.2 Data mining and Spurious Data

Jegadeesh and Titman (1993) first made the momentum effect more widely accepted by
getting their findings published in the Journal of Finance. One common answer from the
traditionalists was that the data was spurious, i.e. that Jegadeesh and Titman had looked at
too short periods, special periods and so on. This critique was answered by Jegadeesh and
Titman (2001), where they repeat their tests with more recent data and with longer periods
of earlier data. Of course, to really ”prove” the momentum effect, one would have to test
all available data from all countries, but the Jegadeesh and Titman tests at least seem to
confirm that there exists or has existed a momentum effect for American stocks.
One case that has been made against the momentum research is that the returns only
reflect recent U.S. history. An attempt at dealing with this claim was made in Rouwenhorst
(1998), where momentum strategies in 12 European countries, including Sweden, between
1985 and 1990 are tested. Rouwenhorst found that the strategies worked in all 12 countries,
and hence that the momentum effect is not restricted to the United States. See Section 2.4
for further discussion.
4
The authors do not investigate the momentum effect explicitly, but that stocks have momentum is an implic-
ation of the fact that they tend to revert.
2.2 A Model of Over- and Underreaction

With this model we will attempt to illustrate, nota bene not prove, the concepts of over- and
underreaction and how those can affect the prices of stocks leading to the momentum-effect.5
The model is thus a theoretical motivation to why our trading model could work in reality.
The model is a so called ”representative agent”-model, i.e. there is only one agent in the
model. We further assume that the investor is risk neutral and that 100% of earnings are
paid out as dividends.6 Hence, the agent’s valuation of the earnings-stream will equal the
price of the stock. This investor is not aware of the true underlying process governing the
earnings process, which is a random walk (i.e. completely random):
Nt = Nt−1 + yt , (4)
where Nt is the earnings in period t and


 y with probability 0.5
yt =

−y with probability 0.5,
is the period t shock to earnings. The earningsprocess is a martingale, since E[Nt+1 |Nt ] =
Nt + E[yt+1 ] = Nt + (0.5)y − (0.5)y = Nt and E(|Nt |) < ∞.7 If the investor were fully
rational, and would have the computational capacity to realize that the earnings-stream is
a random walk, her valuation of the stock would just be
· ¸
Nt+1 Nt+2 Nt
Pt = Et + 2
+ . . . = , (5)
1+δ (1 + δ) δ
where δ is the investor’s discount factor.8 With this valuation, there would be no over- or
underreaction and hence no momentum effect.
The investor in the model at hand is however not fully rational, and instead believes that
the shocks to earnings are governed by either one of two models, either a mean-reverting
model or a model of positive autocorrelation (earnings trend). She will then use whichever
model she thinks governs earnings to forecast dividends and prices. Depending on the shocks
to earnings the investor observes at time t, she will shift between the two models. Both the
two models and the process by which the investor switches between the two are Markov
5
The model is Barberis’, Shleifer’s and Vishny’s model as presented in Shleifer (2000).
6
The assumption of risk neutrality is done for the sake of simplicity.
7
We have used the definition of a martingale found in Djehiche (2000).
8
Equation (5) holds, since Nt is a martingale and hence the expression reduces to
X
∞
1
Nt ,
n=1
(1 + δ)n
which is a geometric series, and can be rewritten as

!
X
∞
n
Nt x −1 , (6)
n=0
1 1
where x = 1+δ
.The sum equals 1−x = 1− 1 1 = 1+δ δ
(see e.g. Adams (1995, p. 530)). Equation (6) thus reduces
h 1+δ i
N N Nt
to Nt ( 1+δ
δ
− 1) = Nt 1δ , so Et 1+δ
t+1
+ (1+δ)
t+2
2 + . . . equals δ .
processes (see Appendix A.6 for an explanation of Markov processes).

Model one9 :
yt+1 = y yt+1 = −y
yt = y πL 1 − πL
yt = −y 1 − πL πL
This will be a mean-reverting model if πL ∈ (0, 0.5).
Model two:
yt+1 = y yt+1 = −y
yt = y πH 1 − πH
yt = −y 1 − πH πH
The shocks according to model two will be positively autocorrelated if πH ∈ (0.5, 1).
We can write the transition matrix for the Markov process governing the switches between
the two models as
st+1 = 1 st+1 = 2
st = 1 1 − λ1 λ1
st = 2 λ2 1 − λ2 ,
where st = i, i = 1, 2 means that at time t, the investor believes that model i governs the
earnings shocks.
As mentioned above, the valuation of the stock would be really simple if the investor knew
the process was a random walk, but now we have to take into account that the investor’s
beliefs regarding the governing process change over time between the mean-reverting model
and the model of positive autocorrelation. According to the strings of news she receives,
where news is defined as the shock to earnings in a specific period, she updates her beliefs
according to Bayes’ rule. A shock is considered as good news if it is positive and as bad
news if it is negative. To make this formal, assume that the investor at time t calculates qt ,
the probability that the shock observed in period t was generated by model 1, using the new
data to update her beliefs from the preceding period, qt−1 . This means that qt = P (st =
1|yt , yt−1 , qt−1 ). The process of updating can be expressed as qt+1 =
((1 − λ1 )qt + λ2 (1 − qt ))P (yt+1 |st+1 = 1, yt )

((1 − λ1 )qt + λ2 (1 − qt ))P (yt+1 |st+1 = 1, yt ) + (λ1 qt + (1 − λ2 )(1 − qt ))P (yt+1 |st+1 = 2, yt )
(7)
This expression follows from the definition of transition probabilities above and the use of
Bayes’ rule. To get a feeling for how this updating works, we have plotted a string of shocks
and the corresponding qt s in Table 1. When the investor sees a string of positive or negative
shocks, she gets more and more certain that the earnings are generated by the positive auto-
correlation model, i.e. qt sinks. When she observes shocks of alternating good and bad news,
she gets more and more convinced that the earnings are generated by model 1, i.e. qt goes
up.
We are now ready to explore how the investor values the stock. As we mentioned above,
a fully rational investor would, according to the EMH, value the stock as Pt = Nδt , but this
investor believes that the earnings are generated by the model-switching scheme described
above. Thus, she will value the stock according to the following formula
Nt
Pt = + yt (p1 − p2 qt ), (8)
δ
9
This is a transition matrix, which means that if the initial state is yt = −y, the next state will be yt+1 = −y
with probability πL etc.
Table 1. Illustration of how the investor changes her beliefs about which model governs the earnings
according to the shocks to earnings she observes, when the shock to earnings is 10. qt is the
probability that she assigns to the earnings being generated by model 1, i.e. the mean-reverting
model. The qt s are rounded to four decimal places.
Time Shock qt
0 10 0.5000
1 10 0.4000
2 10 0.3429
3 -10 0.7318
4 -10 0.5573
5 10 0.8223
6 -10 0.9110
7 10 0.9364
8 -10 0.9433
9 10 0.9451
10 10 0.7436
where p1 = 1δ (γ0T (1 + δ)(I(1 + δ) − Q)−1 Qγ1 ) and p2 = − 1δ (γ0T (1 + δ)(I(1 + δ) − Q)−1 Qγ2 ).
The transpose of a matrix or vector x is written xT . γ0T = (1, −1, 1, −1), γ1T = (0, 0, 1, 0),
γ2T = (1, 0, −1, 0).
 
(1 − λ1 )πL (1 − λ1 )(1 − πL ) λ 2 πL λ2 (1 − πL )
 (1 − λ1 )(1 − πL ) (1 − λ1 )πL λ2 (1 − πL ) λ 2 πL 
Q= 

λ 1 πH λ1 (1 − πH ) (1 − λ2 )πH (1 − λ2 )(1 − πH ) 
λ1 (1 − πH ) λ 1 πH (1 − λ2 )(1 − πH ) (1 − λ2 )πH
The transpose of the matrix Q describes the transitions between the models and the earnings
shocks,
A plot of the earnings and corresponding prices is given in Figure 1 to make the reader
aware of how the model looks. To give an indication of how the ”true” price differs from
the value perceived by the representative investor, we have plotted some values for both the
price process discussed above and the prices if the investor knew the underlying process was
a random walk in Figure 2. This shows the essence of the momentum effect, the over- and
undershooting of price in relation to the ”rational” price.
The difference between the ”true” price and the actual price is represented by the last
term in Equation (8), which can be thought of as the deviation from the fundamental value.
This bias gives rise to the over- and underreaction effect. If the overreaction effect holds,
then, according to our claims above, we would observe a lower return on the stock following
a period of good news, than following a string of bad ones. If, on the other hand, the
underreaction effect holds, we would observe a result in line with Equation (3), i.e. the
return on a company’s stock is higher following an announcement of good news than the
return following bad news. To test these effects within the framework of the presented
model, we simulate the earnings of 50 ”companies” for 6 periods, where one period could be
e.g. a year. A company is defined as a six-period earnings string.
4 4
If overreaction existed in the given model, we should observe that r+ is lower than r− ,
10
in line with Equation (2). This is the observation of Shleifer (2000), but we want to take
the analysis further, by getting an indication of the significance of the results, which is not
done in Shleifer (2000). To test this, we run the described simulation of 50 firms 100 times
in Matlab. This leaves us with a sample of 100 observations.
To check the existence of the overreaction effect we look at the average difference between
4 4
r+ and r− . This is done by constructing a 99% confidence interval for the average differ-
ence. The distribution of the differences is plotted in Figure 3. Knowing that each run is
4
10
We denote the return on stocks of companies with four consecutive years of positive earnings shocks r+ .
Earnings
Price
700
600
500
Price and Earnings
400
300
200
100
0
0 100 200 300 400 500 600
Time
Figure 1. Realizations of earnings over 600 periods and the corresponding prices. To see the
difference between irrational and rational pricing, see Figure 2.
"True" Value of Stock

380 Value Perceived by Irrational Investor
370
360
350
340
Price
330
320
310
300
290
280
0 5 10 15 20 25 30 35 40
Time
Figure 2. The essence of the momentum effect. 40 realizations of the earnings and the corresponding
prices if the investor knows the underlying process is a random walk vs. the prices perceived by the
investor when she does not know the price process. This price pattern can be exploited by a trading
strategy.
independent of the others and assuming that each run is identically distributed, with finite
expectation and variance, we can, by the central limit theorem (see e.g. Gut (1995, pp.
173-174)), construct a confidence interval of the following form:
tn−1,α/2 sx tn−1,α/2 sx
x̄ − √ < µ < x̄ + √ , (9)
n n
where x̄ and sx are the sample mean and standard deviation respectively, n is the sample
size and µ is the mean of the population (see e.g. Newbold (1995)). We find that
P (−4.15 · 10−3 < µ < −4.09 · 10−3 ) = 0.99,
which is consistent with the overreaction effect, since the average return difference is so
clearly below zero.
To check for the existence of the underreaction effect, we examine the average difference
1 1
between r+ and r− . This is done by constructing a 99% confidence interval for the average
difference. The distribution of the differences is plotted in Figure 4. The 99% confidence
1 1
interval for r+ − r− , calculated analogously to above, is
µ ∈ (2.62 · 10−4 , 3.06 · 10−4 )
This is consistent with the formal definition of underreaction, Equation (3) above, in the
sense that it is clearly above zero.
When considering our results, please note that we have used a specific parameter setting
(for π and λ). The model will not produce significant over- and underreaction results for
all possible parameter specifications. For a further discussion of the stability issue, the
interested reader is referred to the original working paper (Barberis and Shleifer, 1997), and
the discussion below.
2.2.1 Conclusions and Discussion

To conclude, we first state three criteria that are generally considered as relevant when
evaluating models of this kind (Hong and Stein, 1999):
• The assumptions underlying the model should be a priori plausible or consistent with
casual observation
• The model should explain existing evidence in a parsimonious and unified way
• The model should make a number of further predictions that can be subject to out-of-
sample testing and that are ultimately validated
The assumptions are very robust, since they are based on the results of research both in finan-
cial economics and in experimental psychology, see e.g. Bondt and Thaler (1985) and Kahne-
man and Tversky (1979). One draw-back though, is that there is only one agent, which means
that the model fails to capture the effect of traders with heterogenous beliefs. This issue has
been addressed in the famous study by Hong and Stein (1999). They also provide a unified
account of over- and underreaction, but with two groups of traders.
An interesting question is how the model would change if the investor were risk averse,
which is assumed in e.g. CAPM. We would then have needed to specify a utility function
for the investor, which would affect her valuation of the stock. This would increase the
arbitrariness of the model, since utility functions are not observable in reality.
The most obvious strength of the model employed is that it is parsimonious and that it
is unified, i.e. it covers both the under- and overreaction effects. A weak side is that it really
25
20
15
10
0
−0.04 −0.03 −0.02 −0.01 0 0.01 0.02 0.03
Figure 3. A histogram describing the distribution of results of the simulation of the overreaction
effect. The mean of the differences is −4.12 · 10−3 and the standard deviation is 9.96 · 10−5 . The
existence of the overreaction effect is significant.
30
25
20
15
10
0
−0.03 −0.02 −0.01 0 0.01 0.02 0.03
Figure 4. Results of the simulation of the underreaction effect. The mean is 2.84 · 10−4 and the
standard deviation 8.55 · 10−5 . Our findings are consistent with the underreaction effect.
does not make any further predictions, but this issue is not relevant for our thesis. There are
some points which the model above does not address, e.g. how arbitrageurs would affect the
results. One may argue that the addition of arbitrageurs would eliminate the effects. This
is not necessarily the case, since arbitrage is quite risky in this setting. No one can perfectly
predict when and how prices will reach the fundamental value. Trend following arbitrage
trading can result in a self-feeding bubble, that makes prices deviate even further from the
fundamental value (de Long et al., 1990). This accentuates stock momentum even further.
Increased volatility is not only beneficial however, since exact timing will most probably be
harder.
Regarding the results of the theoretical model above, we have seen that the representat-
iveness heuristic and the conservatism bias could be explanations for the over- and underre-
action effects, which together drive the momentum effect. As can be seen below, this claim
is also supported by empirical evidence, see e.g. Table 2. Underreaction and overreaction
cause over- and undershooting of prices, in the sense that prices of a stock keep irrationally
trending up after the release of good news, and vice versa when the news is bad. The model
depicted above implies that a typical investor can be seen as using the representativeness
heuristic, i.e. she sees systematic patterns in sequences that are in fact completely randomly
generated. This explains the momentum effect according to this model.
But, having said this, one can still question the exact parameter specifications of this
model, and the circumstances under which it produces economically significant results. We
e-mailed Professor Shleifer and asked him if he could motivate these parameter settings (e.g.
the transition probabilities) from an empirical standpoint. He replied and commented that
he could not, but that we indeed had an important point in questioning this and that he
also very much would like to know the answer to this question and possibly weak link in his
model. Once again, theoretical reasoning needs to be complemented with empirical facts in
order to say something meaningful about reality.
2.3 Theoretical Motivation for the Stop-Loss Rule

This section offers a theoretical motivation for the use of the stop-loss rule, which is often
incorporated in trading models. In its most simple form, the stop-loss rule means that one
liquidates a losing position when its price exceeds or drops below a certain limit, depending
on if it is long or short. Its use is motivated by the fact that one gains by taking small rather
than large bets when one is subject to “bankruptcy costs”. Assume a situation where one
does not limit one’s bets. A well-known such situation is the martingale betting strategy,
a version of which we will go through briefly here.11 The strategy is called the martingale
betting strategy since the sequence Sn defined below is a martingale with respect to itself.
Assume that a game is played where the player bets on the result of a toin coss. The player
wins her bet if the coin comes up a head and loses if it is a tail. The strategy is simple:
double the bet until the preferred side of the coin comes up, which it is bound to do, say at
time T . The profit of this strategy every time one wins is
2T − (1 + 2 + 4 + · · · + 2T −1 ) = 1 (10)
The first term on the left-hand-side of Equation (10) is the player’s gain at time T and the
second term is the money she has lost up to and including time T − 1. Let Sn denote the
11
This example is adapted from Djehiche (2000).
accumulated gain up to the nth toss, S0 = 0. Then Sn evolves according to


 Sn − 2n with probability 0.5
Sn+1 =

Sn + 2n with probability 0.5
The probability that the game stops after the nth toss, i.e. that the player wins, is given by
11 1
P (T = n) = ... = n,
22 2
for n ≥ 1, since there is a probability of a half that the coin comes up head. This means
that the mean loss of the gambler tends to the following limit as n grows large:
X∞
1
(1 + 2 + · · · + 2n−1 ) = ∞,
n=1
2n
where the sum in parentheses is the same as the second term of the left-hand-side of Equa-
tion (10). This implies that the initial capital of the gambler, or investor, needs to be infinite,
if she intends to pursue her martingale betting strategy. Otherwise, after a string of unfa-
vourable events, the gambler will go bankrupt (this is the strategy pursued by casinos when
limiting the maximum size of bets). The essence of the stop-loss rule is precisely to minimize
the risk of losing large ”bets”, so therefore the managers of funds using the stop-loss rule will
not be needing unlimited funds, or at least not as much access to capital.
2.4 Empirical Evidence Using CAR

Now, we turn our attention to the empirical evidence of primarily the momentum effect.
These general facts will later be used as guidance when we test different trading rules in our
model. Also, we will later emulate the simplistic methodology (as in only using the price
variable, no transaction costs, static rebalancing etc) used here, to compare it with our own
model as stated in the purpose of this thesis.
The research cited in the following text often takes the form of a statistical test for
abnormal returns in a market neutral portfolio that is rebalanced periodically a few times a
year in accordance with a ranking scheme based solely on price changes. It is thus a static
version of the more dynamic model that we try to build.
2.4.1 The Momentum Effect

As we have seen in Section 2.1 above, the momentum effect is basically return continuation
in the medium term in both negative and positive direction, followed by mean reversion over
the longer term. The reason: economically irrational investors.
The basic methodology of the momentum studies cited in Table 2 is, as stated above,
to rank stocks after a ”ranking period”. Normally, the 10-20% best and worst performers
form a long and short portfolio, respectively. This combined portfolio is a close to zero β
investment that in theory should yield 0% in return (save transaction costs and interest on
short positions). Then, one looks at a ”test period” directly after the ranking period and
test statistically if the strategy yields any abnormal return above 0%, in which case the
momentum effect would be present. The results are presented as a Cumulative Abnormal
Return (CAR) in the test period. Different lengths of the ranking and test periods are then
studied. Transaction costs are excluded in the cited research, leading to an overstatement
of achievable returns. Furthermore, these strategies are very simplistic, which also affects
performance, as we will see later. And CAR, accompanied with a t-statistic, does definitely
not give the whole picture necessary to correctly assess the true potential (in a negative or
positive sense) of the momentum effect.
Table 2. A brief summary of empirical research on the momentum effect.
Paper Data Sample Results Other Conclusions

(Period)
L.K.C. Chan NYSE, Amex, Nasdaq Significant momentum effect, both Continuing positive and negative trends in the
and Lakonishok (1973-1993 and 1994- for ranking based on 6-month re- months following portfolio formation are attrib-
(1999) 1998) turn and changes in earnings ex- uted to ”conservatism bias” among investors, be-
pectations. ing slow to adjust to new information.
D. Schiereck and Frankfurt Stock Ex- Significant momentum effect. See Beta, risk or firm size does not account for mo-
Weber (1999) change (1961-1991) Table 3. mentum, but behavioral theory does provide a
plausible explanation.
The momentum effect is similar to what is found
in US data despite differences in markets.
Predictable reversals of earnings and price for the
winning and losing portfolios.
M. Finn and S&P 500 (1983-1998) Significant momentum effect. A Most of the profit from a long/short strategy us-
Kling (1999) long/short portfolio would yield ing large capitalized stocks comes from the short
7.8% with a tracking error of 4.8% side, due to high analyst coverage and limited
and beta close to zero, with respect short selling by large investors in this universe.
to the S&P 500.
The momentum effect decreases with company
size.
If one controls for B/M and size, economic sector
is of a less importance, probably sample specific
however.
Jegadeesh and NYSE, Amex (1965- Reaffirms support for previous Significant mean reversion of momentum in the
Titman (2001) 1997) studies that a long/short portfolio 13-60 month period following portfolio formation.
can earn approximately 1% excess The momentum effect can not be explained
return per month. by cross-sectional dispersion of expected return,
only by behavioral theories.
Swaminathan and NYSE, Amex (1965- Significant momentum effect. Trading volume works as a good predictor.
Lee (1999) 1995) The best shorting candidates are high volume low
momentum with a ranking period of six months.
The best buy candidates overperformed less than
the sell candidates underperformed.
Daniel and Tit- NYSE, Amex, Nasdaq Significant momentum effect. 13% Stronger momentum effect for low B/M stocks,
man (1999) (1963-1997) yearly return if one takes a long po- since they are more susceptible to investor senti-
sition (20% most extreme) in high ment.
B/M high momentum and shorted Rejection of adaptive market efficiency in favor
the low B/M low momentum stocks, of behavioral explanations.
t-statistic 5.66.
O’Neal (2000) S&P (1989-1999) Much of the observed momentum in individual

stocks are attributed to industry momentum.
In the study by D. Schiereck and Weber (1999), who studied the German market, the
results of their investigation of the momentum effect are compared with the classic study
by Jegadeesh and Titman (1993) using US data. Table 3 shows that the two studies have
similar return patterns and that there are significant momentum effects when the ranking
period is more than or equal to 3 months. The CAR increases generally at a decreasing rate
as the test period reaches 6 to 12 months where mean reversion sets in. Note that the CAR
is measured as a total up until the end of the test period, and not as a monthly return.
The latest study on the momentum effect for Swedish data was done as a Master Thesis
at Stockholm School of Economics in 2000, presented in Table 4. The CAR numbers are
here presented on a monthly basis, not total. They also found significant results comparable
to previous studies. However, during 1992 when Sweden devaluated its currency, the losing
Table 3. The momentum effect in Germany, compared with Jegadeesh-Titman’s results for USA,
1961-1990. 20 W-L means that the 20 winning and 20 losing stocks in the ranking period form
the market neutral long/short portfolio. Sample size: 357 most traded stocks. Reproduced
from D. Schiereck and Weber (1999)
Strategy CARrank period CAR3 CAR6 CAR9 CAR12
One-month rank period

20 W 1961-90 13.18% 0.00% 0.35% 0.02% 0.78%
(t-statistics) (0.02) (2.38) (0.17) (5.64)
20 L 1961-90 -10.43 -0.20 -0.15 -0.25 -0.70
(-1.24) (-1.03) (-1.70) (-4.65)
20 W-L 1961-90 23.61 0.20 0.50 0.28 1.49
(0.78) (2.14) (1.24) (6.35)
20 W-L 1961-70 22.06 -0.62 0.73 0.22 0.96
20 W-L 1971-80 21.37 0.82 0.74 0.38 1.33
20 W-L 1981-90 27.41 0.41 0.04 0.24 2.17
Three-month rank period
20 W 1961-90 20.76 0.31 1.46 2.10 3.47
(0.99) (3.10) (3.32) (4.78)
20 L 1961-90 -19.58 0.38 -0.07 -0.80 -2.06
(1.06) (-0.18) (-1.40) (-3.05)
20 W-L 1961-90 40.35 -0.07 1.53 2.90 5.52
(-0.15) (2.45) (3.40) (5.57)
20 W-L 1961-70 39.01 -1.35 0.02 1.46 3.59
20 W-L 1971-80 36.32 0.13 1.17 1.42 2.17
20 W-L 1981-90 46.73 1.00 3.40 5.81 10.81
20 W-L 1965-89 Na -0.03 1.56 2.61 5.40
Jegadeesh-Titman Na 0.96 1.74 5.40 8.28
Six-month rank period
20 W 1961-90 30.07 0.28 1.95 3.12 4.15
(0.46) (2.24) (2.95) (3.66)
20 L 1961-90 -29.44 0.37 -1.51 -2.73 -3.92
(0.67) (-2.09) (-2.78) (-3.34)
20 W-L 1961-90 59.51 -0.09 3.46 5.84 8.07
(-0.11) (3.06) (4.06) (4.95)
20 W-L 1961-70 55.90 -0.39 0.05 0.74 2.26
20 W-L 1971-80 53.38 0.88 4.99 8.31 8.83
20 W-L 1981-90 69.31 -0.77 5.33 8.48 13.13
20 W-L 1965-89 Na 0.09 4.32 6.75 8.76
Twelve-month rank period
20 W 1961-90 47.05 1.77 3.06 3.06 4.14
(1.88) (3.27) (2.33) (2.59)
20 L 1961-90 -45.90 0.98 -1.84 -0.83 -1.07
(0.83) (-1.57) (-0.52) (-0.47)
20 W-L 1961-90 92.95 0.80 4.90 3.89 5.21
(0.53) (3.26) (1.88) (1.87)
20 W-L 1961-70 81.12 -2.01 0.96 0.27 1.06
20 W-L 1971-80 83.71 4.83 8.85 8.55 5.31
20 W-L 1981-90 114.04 -0.43 4.90 2.85 9.24
20 W-L 1965-89 Na 1.11 5.40 4.95 3.52
portfolio with negative momentum suddenly outperformed the winning portfolio, resulting
in a large loss for the long/short portfolio. The authors therefore eliminated one year of
data to evaluate the impact of the devaluation, and one can clearly see the improvement
in performance. Eliminating bad data is of course a dubious way of proving a strategy
profitable, but it highlights an important aspect of our strategy: risk control. If one does
not take into consideration the composition of the long and short positions, one will surely
end up with a lot of exposure to various risk factors. A long stream of steady profits can
disappear literally overnight.
2.4.2 The Role of Volume, Fundamental and Sentiment Variables

The most important variable when modeling the momentum effect is of course the price
itself, as was evident in Shleifer’s model. It incorporates investors’ beliefs and will thus also
indicate whether a specific stock is ”neglected”, with negative momentum, or a ”glamour”
stock with positive momentum. But there are other variables to consider.
Table 4. The momentum effect in Sweden 1980-1999. Stockholm Stock Exchange, between 118
and 233 stocks. The Table is reproduced from Söderström (2000).
Strategy CAR3 CAR3 CAR6 CAR6 CAR9 CAR9 CAR12 CAR12
1 year after 1 year after 1 year after 1 year after
devaluation devaluation devaluation devaluation
eliminated eliminated eliminated eliminated
Three-month rank period
10% W-L 1980-1999, monthly return -0.37% -0.05% 0.22% 0.63% 0.25% 0.53% 0.16% 0.56%
(t-statistics) (-0.83) (-0.12) (0.61) (2.40) (0.92) (2.12) (0.53) (2.54)
monthly standard deviation 3.90% 3.61% 2.80% 2.22% 2.51% 2.15% 2.77% 1.95%
Six-month rank period
10% W-L 1980-1999, monthly return 0.37 0.85 0.66 0.99 0.47 0.87 0.19 0.68
(t-statistics) (0.70) (1.86) (1.99) (3.30) (1.49) (3.43) (0.59) (3.11)
monthly standard deviation 2.27 1.90 2.96 2.52 2.90 2.16 2.93 1.90
Nine-month rank period
(t-statistics) (1.25) (2.03) (2.13) (4.19) (1.52) (4.53) (0.34) (3.78)
Twelve-month rank period
(t-statistics) (0.68) (2.36) (1.03) (4.26) (0.76) (4.19) (0.05) (3.12)
The volume variable can be viewed in two ways. Either as a proxy for analyst coverage,
where high volume means bad momentum characteristics (high analyst coverage is bad for
long positions, but can be good for short positions since analysts tend to push valuations to
their upper limits), since the price will be less exposed to sentiment changes and irrational
pricing. Or as a proxy for strong positive feedback trading, if the volume exhibits an increas-
ing trend during price increases or decreases. The volume variable is therefore an example
of a variable that can be used in many ways to find the optimal composition of a momentum
portfolio.
Fundamental variables can be effectively used for value stocks, where a lot of the value-
creating process is reflected in the quarterly reports. Piotroski (2000) constructed a market
neutral portfolio using US data on high B/M firms between 1976 and 1996. He ranked the
stocks quarterly using 9 measurements of fundamental strength and achieved a 23% annual
return. The problem with this approach in our case is that we have not got quarterly data
and that we focus on the price variable where growth stocks are more prone to trend. We
will however use the P/E, M/B and market cap variables as proxies for analyst coverage and
growth characteristics. Furthermore, we have data on analysts’ beliefs about cash flow one
year ahead, which in essence is a combination of a fundamental and a sentiment variable.
2.4.3 Conclusions
Reviewing this section, we find empirical support for the existence of the momentum effect
and some interesting results on more specific characteristics of the momentum effect. We
will not in this thesis discuss exactly how we translate these specific facts into algorithms;
a non-mathematical summary is however given in Table 12 in Section 5. Not all of the
variables in our model are even directly referenced in this section, but rather used as proxies
for underlying phenomena.
One specific issue needs commenting, though. In the simplistic strategy that we emulate
later in this thesis, we use a 6-month ranking period and a 4-month holding period. These
parameter values yield approximately the best results, judging both from the empirical res-
ults presented here and our own simulations. In the more advanced model that we have
constructed ourselves, one of the rules also has these settings, although somewhat modified
by giving more weight to price changes the last week.
3 AN OVERVIEW OF OUR MODEL AND ITS CONTEXT 19
3 An Overview of Our Model and Its Context

3.1 Different Hedge Fund Strategies
We will here give a brief overview of the hedge fund market in order to better understand
the context of the kind of model we try to build and the strategy behind it.
The term ”hedge fund” has evolved over time to include a multitude of skill-based in-
vestment strategies with a broad range of risk and return objectives. The common element
among these strategies is the use of investment and risk management skills to seek positive
returns regardless of market direction. A hedge fund can use investment techniques that
include derivatives, leverage and arbitrage in any market in order to generate specific risk-
return profiles. The largest incentive to invest in a hedge fund is the diversification benefit
for the investor’s portfolio due to the attractive risk-return profiles of hedge funds, often due
to the low correlation with stocks, the largest asset class.
Below are some quick facts about the hedge fund industry:
Table 5. Hedge fund history at a glance. Reproduced from Säfvenblad (2001).
History:
1949 AW Jones formed the first hedge fund. He used long/short equity positions and
leverage, and charged a performance-based incentive fee.
1966 Fortune ran an article on Jones’ fund, and public interest led to many new hedge
funds.
1970 Hedge funds entered their first crisis as equity markets fell. Industry studies es-
timated that hedge fund assets under management fell by 70%.
1986 Institutional Investor ran an article on the Tiger fund which renewed interest in
hedge funds.
1992 George Soros speculates against several European currencies with his Quantum
fund and makes a fortune.
1997 Hedge funds blamed for triggering the Asian currency crisis
2000 6000 hedge funds, $ 450 billion in assets
Asset allocation within hedge fund universe in 2000:

• Long/Short Equity 40%
• Event Driven 15%
• Managed Futures 14%
• Global Macro 14%
• Fixed Income Arbitrage 7%
• Convertible Arbitrage 4%
• Other 6%
The classification of different hedge funds is not entirely straightforward since some funds
may use a combination of strategies, if they disclose anything at all. Strategies can be
categorized in many ways. Using the definitions by Morgan Stanley as in Tables 6 and 7
below, we can see that our return generating technique is categorized as a statistical arbitrage
investment process.
In Table 8, we see how three data-providers have divided the hedge fund universe into
sub-categories. The performance data gives an indication of the similarities between groups
of funds and what their characteristics are.
Table 6. The investment process main and sub-category designations for hedge funds. The cat-
egories are Morgan Stanley’s.
Hedge Fund Strategies
• Directional Trading
– Discretionary Trading
– Strategic Allocation
– Systematic Trading
• Relative Value
– Convergence Arbitrage
– Merger Arbitrage
– Statistical Arbitrage
• Specialist Credit
– Distressed Securities
– Positive Carry
– Private Placements
• Stock Selection
– Long Bias
– No Bias
– Short Bias
– Variable Bias
Table 7. Definition of Statistical Arbitrage.
An approach is defined to be Statistical Ar-

bitrage if . . .
• positions are entered into on the basis of

systematic models designed to find oppor-
tunities where the relative value of two or
more assets is currently different from a
theoretically or quantitatively predicted
value, and
• manager intervention is limited to redu-
cing or eliminating the trades entered into
in less than 25% of cases, and
• the portfolio is comprised solely of indi-
vidual securities within one particular as-
set class, and
• the portfolio net exposure is never more
than 20% long or 20% short, and
• the approach does not conform to Con-
vergence Arbitrage or Merger Arbitrage
3.2 Our Model – an Optimization Problem

There are basically three important aspects of choice that define the characteristics of a
strategy. Our choices are displayed in Table 9.
In essence, the problem of building a statistical arbitrage model that yields an optimal
risk-return profile is an optimization problem. The input data makes up the independent
variables, the model itself is the function, and some risk-return measure is the dependent
target variable. Certainly, this is not the ordinary linear maximization problem that is
analytic and algebraically solvable. The function is unknown, the full data-set is unknown
and the data is stochastic. Mathematically, we have a multi-variable problem with many
local maxima that are not stable over time.
The idea is thus, in a semi-deductive way, to try to localize approximate areas in the
Table 8. Performance Jan. 1990 - Dec. 1999 from three different hedge fund data providers.
Reproduced from Könberg and Lindberg (2000).
Average Annual Sharpe
Annual Standard Ratio
Benchmark Indexes Return Deviation
S&P 500 Composite 18.90% 13.15% 1.05
MSCI World US 12.99% 14.75% 0.53
Lehman Government 7.46% 4.10% 0.57
Lehman Corporate 8.47% 4.80% 0.70
Lehman Govt./Corp. 7.70% 4.25% 0.61
GSCI Commodity 6.55% 18.02% 0.08
Salomon Bros. Treasury 7.74% 4.57% 0.57
Salomon Bros. Govt. Corporate 7.87% 4.58% 0.60
Salomon Bros. Wgbi US All Maturities 7.96% 6.50% 0.44
US Treasury Constant Maturities 3 month 5.12%
Hedge Fund Research Indexes
Convertible Arbitrage 11.57% 3.50% 1.84
Distressed Securities 16.83% 6.57% 1.78
Emerging Mkt. Asia 14.45% 14.42% 0.65
Equity Hedge 24.80% 8.58% 2.29
Equity Market Neutral 11.38% 3.16% 1.98
Equity Non - Hedge 23.25% 13.48% 1.35
Event-Driven 17.59% 6.68% 1.87
Fixed Income Total 8.88% 4.00% 0.94
Fixed Income : Arbitrage 9.54% 4.95% 0.89
Fixed Income : Convertible Bonds 17.53% 9.89% 1.26
Fixed Income : High Yield 11.22% 7.28% 0.84
Fixed Income : Mortage-Backed 11.14% 4.63% 1.30
Macro 20.99% 9.24% 1.72
Market Timing 17.75% 8.17% 1.55
Merger Arbitrage 13.02% 4.67% 1.69
Relative Value Arbitrage 14.39% 4.12% 2.25
Short Selling 0.38% 19.77% -0.24
Statistical Arbitrage 11.93% 3.73% 1.82
Fund of Funds 12.78% 6.06% 1.26
MAR Hedge Fund Indexes
Event-Driven 13.85% 4.89% 1.79
Distressed Securities 18.23% 8.43% 1.56
Risk Arbitrage 20.23% 17.96% 0.84
Global Emerging 16.85% 17.51% 0.67
Global Established 18.95% 9.32% 1.48
Macro 15.59% 7.33% 1.43
Market Neutral 11.92% 2.60% 2.62
Market Neutral: Arbitrage 14.97% 6.33% 1.56
Market Neutral: Long/Short 10.69% 1.61% 3.46
Short Sellers 0.51% 17.47% -0.26
Fund of Funds 11.26% 4.60% 1.33
Diversified 11.83% 4.96% 1.35
Niche 10.68% 5.60% 0.99
EACM Hedge Fund Indexes
EACM100 15.62% 4.25% 2.47
Relative Value 10.47% 3.42% 1.56
Long/short Equity 9.72% 3.00% 1.53
ConvHedge 10.50% 4.43% 1.21
BondHedge 7.65% 4.53% 0.56
Rotational 14.09% 6.98% 1.29
Event 13.47% 5.37% 1.55
Arbitrage 10.21% 6.35% 0.80
Bank 15.56% 6.95% 1.50
Multi 14.71% 5.35% 1.79
EQ HEDG 22.31% 9.80% 1.75
DomLong 23.48% 13.15% 1.40
DomOpp 21.77% 9.77% 1.70
Gl/Int 21.69% 11.54% 1.44
GLOB AA 19.84% 10.63% 1.39
Discretionary 18.29% 9.92% 1.33
Systematic 21.41% 17.28% 0.94
Short -0.93% 22.32% -0.27
domain of the function and variables where one is likely to find good results, and then do
a rough optimization with the computer’s help. By avoiding to search for maxima in areas
where there are no theoretical or empirical evidence pointing in that direction, one also averts
data-mining. Only stable maxima are of interest here since they are likely to persist in the
future and yield profitable trading opportunities.
This is the overarching principle that lies behind what is described in Section 5 (Meth-
Table 9. Our strategy.
Aspect of Choice Our Choice Justification
Return Generating Tech- Statistical Arbitrage Strategies based on fundamental analysis, ad-
nique vanced macro-models or experience within an in-
dustry are not interesting since we have not got
the necessary competence. What remain are ba-
sically the more technically inclined strategies,
where programming, statistics and knowledge of
the anomalies of the markets are the critical
knowledge factors.
Investment Style Market Neutral Trading with a ho- Given our statistical arbitrage technique, a mar-
rizon of between a few weeks and a ket neutral approach is suitable because it isol-
few months ates the effect of the arbitrage, and the link
with academic papers on behavioral phenomena
is clear. Furthermore, the important momentum
effect that we want to exploit appears within
the chosen trading horizon. If this horizon is
too short, transaction costs will be prohibitively
large, and if it is too long, prices mean revert
and the effect disappears. One could enhance
the model with varying betas on different factors
(making directional bets), but we limit ourselves
here.
Market Swedish Stocks Equity is the most accessible asset class when it
comes to data-availability and literature on arbit-
rage. It also has the most interesting prospects
in terms of risk-return profile. The exclusion of
derivatives is a limitation from our side. Swedish
data is used for limitation and availability reas-
ons.
odology – Signal Generation and Risk Control).
3.3 Technical Specification of the Computer Program

The computer program, which took us 6 months to develop, functions as a simulation plat-
form. With it we can, using data stored in the Microsoft Access database format, simulate
virtually any trading strategy with a few adjustments in the program. We have built in func-
tions to automatically analyze any interesting statistical property of a user-specified model
executed in the program.
The programming environment employed is Sun’s Java. Among the reasons for this
choice are Java’s relative robustness, its scalability, its position among financial computer
systems in general, its object oriented nature, and its platform independence. The choice of
programming system is not critical, but it is convenient to be able to use it under different
operating systems (Windows, Linux, MacOS and so on). Java 1.4 is the latest version and
therefore probably the better choice when it comes to speed of execution and richness in
features, but the larger installed base of software made Java 1.3 the final choice.
This section requires some understanding of object oriented design, but it should be
enough to note that a program written in a language such as Java or C++ consists of
several classes with different duties and relations among themselves. The main goals in the
development phase were robustness, ease-of-use, speed, and scalability. The program is based
on a scalable plugin architecture, where the plugins are different sets of rules (see Section 5)
that are to be applied to the underlying data (such as price or volume). Implementationwise,
this means that each set of rules is its own class. One minor goal has been to minimize the
coding effort needed to add a new rule. In the current version of the program, it is enough to
write one main function for each plugin and some small methods for determining what kind
of data the plugin needs. Without going in too deep on object oriented design, the rest of the
Figure 5. Conceptual view of the computer system.
functions reside in the superclass. The conceptual design of the system is depicted in Figure 5.
In the diagram, one can view the filter (abstraction for trading rule) MasterFilter and its
relationship to the rest of the system, particularly the core (Ingenium). The collaboration
and hierarchy of different classes are illustrated using arrows. As is hopefully evident, the
modularity of the system makes it an easy task to write a new filter.
A part of the graphical user interface of the system is depicted in Figure 9 in Ap-
pendix A.3. Ease-of-use has been the goal in this area, as well as clarity in design. Ro-
bustness and numerical stability in calculations have been checked using Microsoft Excel
XP. In Appendix A.4, we display an example of what kind of data the simulation platform
can generate in a day of trading.
The source code amounts to more than one megabyte of sheer text, which is approximately
equivalent to 220 pages of text (assuming 4500 characters per page). 270 classes have been
produced from the code.
4 DATA 24
4 Data
The data was electronically retrieved from the international database Datastream and the
Swedish Trust.
4.1 Description of the Data

We have used closing prices for the most liquid Swedish stocks. The stocks are all from
OM’s liquidity classes A and B, and have been adjusted for splits and dividends.12 The
Swedish stock prices were retrieved from Trust. Moreover, data on traded volume, market
value, market-to-book and predicted cash flow for the Swedish stocks were retrieved from
Datastream. See further Tables 10 and 11.
Table 10. Number of companies for which there exist various fundamental data at the beginning
and end of the test period.
M/B Market Value Volume Predicted Cash Flow
95-06-30 69 91 98 65
01-11-06 111 155 150 143
The stocks have been divided into 10 different sectors, based on ”Veckans Affärers Indus-
triindex”, see Table 15 for specification.13 To get a reasonable number of foreign stocks in
each respective market, we set the market capitalization limit to 4.9 billion SKr. The non-
Swedish stocks were downloaded from Datastream. The reason for looking at non-Swedish
stocks is to make out-of-sample tests to ensure that the effects we are studying are not isol-
ated to the Swedish market. Further details can be seen in Table 11. For a list of the stocks
Table 11. The stocks used in the model.

Country Number of Stocks Time Period
France 110-181 900101 - 011106
Germany 110-167 900101 - 011106
Sweden 101-175 950101 - 011106
U K 192-202 900101 - 011106
used please see Table 23 in Appendix A.5.

Apart from the data on stocks, daily quotes of the Swedish one-year rates from 1990-
01-02 through 2001-12-21 have been used. For the calculations of β-values, ”Affärsvärldens
Generalindex” ranging from 1979-12-31 – 2001-11-14 has been used.
Missing data has been approximated as a weighted average of the two closest neighbors.
12
The criteria for the A-group are that the stocks should be traded every day, have an average turnover exceeding
MSkr 4 and a spread under 1%. The B-class of stocks requires that the proportion of trading days exceeds 50%,
that the average daily turnover exceeds Skr 200 000 and that the spread is below 5%. From these classes we have
taken the 175 most liquid stocks.
13
The data is from 1982-09-13 - 2001-11-05 and was downloaded from EcoWin.
5 METHODOLOGY – SIGNAL GENERATION AND RISK CONTROL 25
5 Methodology – Signal Generation and Risk Control

We will in this chapter give a non-technical and non-detailed description of how we have
constructed the algorithms of our trading model (also called default trading model), which is
based on Swedish data and incorporates all trading rules. For the non-Swedish markets, we
have only got the price variable and hence only used a subset of the rules that are applied
to the Swedish market.
A stock in the portfolio can be dropped for two reasons; it hits the stop-loss limit, or it
reaches the end of the holding period when the momentum effect is believed to fade away.
This means that the stocks in the portfolio quickly will move out of phase. In 95% of the days
in our model (see Section 6) there are 0 - 3 trades (sell one & buy another) executed, with
an average of one stock traded every third day. Since we always keep 10 stocks in the long
and 10 stocks in the short portfolio, a dropped stock needs to be replaced immediately (we
have tested the effects of delayed execution in our robustness tests in Appendix A.1.2). And
every time a stock is bought (or shorted), the program needs to evaluate all stocks available
and rank them according to a scheme and then pick a stock from the top (or bottom if it is
a short stock that needs to be added) if it fits the portfolio risk criteria.
We have thus divided this process into two steps. First, the model uses the data and
different algorithms to rank all the stocks according to expected strength and direction in
future momentum, and a ”master list” is created. We have called this step Signal Generation.
Next, we let the program go through the top candidates. For example, if we need a long
(short) stock, the program checks the first 4 stocks (the last 4 stocks) on the master list for
certain properties. The one that best meets our risk criteria is chosen. We have called this
step Risk Control.
5.1 Signal Generation

The purpose of this part of the model is to produce an optimal master list of momentum
stocks. In Sections 2.4 and 2.1 we have found a few guidelines on where to look for good
momentum stocks, which are illustrated in Table 12. For these 5 dimensions, we have
Table 12. Our momentum-proxies.
Characteristics of strong momentum Our Proxy

stocks
Large increase/decrease in price last 6 Stock price 6 months back compared with
months, momentum persists for 4 months today, with an extra weight to the last week
to sort out declining stocks: 20% weight to
the last week and 80% weight to the rest.
Low analyst coverage (long portfolio)/ High Low market cap for long, high market cap
analyst coverage (short portfolio) for short portfolio.
More susceptible to sentiment changes, i.e. High M/B

growth stocks
Positive/negative sentiment in trading Large changes in price during increase in

volume
Fundamental strength/weakness Posit- Large changes in analyst expectations of

ive/negative sentiment of analysts Cash Flow one year ahead
constructed mathematical rules to rank the stocks in accordance with their expected future
momentum. The most important variable is as stated before price, i.e. the rule that ranks
stocks by performance the last 6 months and keeps them for 4 months (although we have
modified this rule somewhat by giving extra weight to the last week). This rule is thus
almost the same as in the simplistic strategy. But by combining this simple rule with others
we have constructed a more advanced and hopefully more effective strategy.
5.2 Risk Control

An investor wants, given the return, as low volatility and low correlation with other assets
as possible. The correlation with the market (or beta) measures this exposure assuming
that the investor holds a diversified portfolio. In a market neutral portfolio such as ours,
the market risk will be rather low (see Section 6). But there are other risks that are not
correlated with the market. These risks stem from the fact that the long and short portfolios
may contain stocks that belong to groups of stocks that are negatively correlated. A decrease
in the price of the long portfolio and a simultaneous increase in the stock prices of the short
portfolio will result in a large loss. It is therefore important to neutralize these risks.
The most obvious risk is that the long and short portfolios end up with a large exposure to
different industries that tend to move in the opposite direction of each other. This frequently
happens when one ranks stocks on price changes alone, where one industry is located on the
top of the chart and another industry close to the bottom. If there is a turnaround in the
economy, both the long and short portfolio will start to lose money at the same time.
Other categorizations of stocks that tend to move in cycles are growth versus value stocks
and large cap versus small cap stocks. The growth category is overweight in both the long
and short portfolios and will as such not pose a large threat to the stability of the portfolio.
The relative small overweight of large caps in the short and small caps in the long portfolio
also does not seem to create a systematic risk exposure of great importance (see correlations
with industry index below in Section 6.1), in line with our observations in Section 2.1.
What we basically have then is a problem of covariance within the portfolio. Our main
solution is that we test the best 4 candidates for inclusion in the portfolio by calculating the
sum of the variance-covariance matrix, one at a time, and then pick the stock that results
in the lowest portfolio risk. By testing 4 candidates, one will increase the chance of finding
a stock that has strong momentum but also a beneficial volatility structure.
We have three more categories of risk control; the stop-loss rule, a low price cut-off and an
indicator of extreme valuations. We do not trade when the stock price is under 3 SKr. These
stocks often move in large discrete steps in an unpredictable manner. The stop-loss rule is
set to 20% of the maximum value during the holding period. Interpreting the 20% stop-
loss as a contrarian signal motivates this rule, since expected momentum profits diminish
with contradictory signals (since the basis for momentum is return continuation). When a
stock hits the stop-loss limit, it is banned from entering the portfolio for 15 trading days (3
weeks). And when 4 stop-losses are executed within one month and industry (as defined by
correlation with AFGX indices 6 months back), the whole industry is banned from entering
the portfolio for one month and an opposite position is taken within the industry relative
to the original position. This rule remedies situations where an industry has had a strong
momentum for a long time and suddenly turns around. This industry will be overweight
in the long (or short) portfolio, and some signals will still indicate momentum in the old
direction before the turnaround.
The indicator for extreme valuations is also targeted to challenge the problem of turn-
arounds. The rule bans industry sectors on the long (short) side where the M/B has doubled
(halved) the last year for more than 4 stocks in the sector. When valuations deviate too
much from fundamental value, prices start to converge again. In the words of Shiller (2000),
irrational exuberance can not hold forever.
5.3 Assumptions in Our Simulations

In all of our simulations, we let the program use three separate accounts to make the simu-
lations as realistic as possible. These are:
1. The Long Portfolio
2. The Short Portfolio
3. The Transaction Account (TA)
We start out with an equal value of the long and short position (consisting of 10 long, 10
short stocks respectively) and a blank Transaction Account, where the transaction costs for
taking the initial positions are subtracted at the end of the first day. When a long position
is sold it is booked in the TA, and when a long position is bought, money is withdrawn from
the TA. All the transaction costs for both the long and short portfolios are booked in the
TA. Except for the transaction costs, the short portfolio (which is a zero net investment) will
affect the TA every time one closes a position, leading to a net profit or loss. In essence, the
structure of the accounts and the transactions between them should mimic the real world
situation.
The following definitions have been made:
• Value of Total Portfolio(t) = [Value of Long Position(t)]+[Value of Net Short Position(t)]+
[Value of Transaction Account(t)]
• Total Return(t) = log[Value of Total Portfolio(t)/Value of Total Portfolio(t − 1)]
5.3.1 Basic Assumptions of Our Trading Model

These are all true unless otherwise stated.
• We have 10 short and 10 long positions at all times. This is of course a variable that
could be changed in a better model. But it is a reasonable limitation from our side
considering that our sample of less than 175 stocks would cause less profitability on
the margin if one includes many more stocks, and that fewer stocks would make the
portfolio too undiversified.
• Our default trading model is run from 1995-06-30 through 2001-11-06. We had to work
with this limitation due to our restricted access to good fundamental data.
• Transaction costs are set to 0.3% for both long and short trades. This should not be
an understatement of the costs for an investor with more than 1 million SKr to invest.
These figures are from Germer (2002).
• When a new position is entered, the basic rule is to make it 10% of the total value of
the long and short portfolios, respectively. This will keep the portfolios from deviating
too much from equal weighting (ten long and ten short positions).
• The Transaction Account, which is initially set to zero, will increase as profitable short
positions are closed and decrease with transaction costs and unprofitable short posi-
tions. We have put two limits on the size of this account; the lower limit being -3%
of Total Portfolio Value, and a higher limit of +3%. The reason is that a too large
negative balance would yield interest rates, and that a too large positive balance would
not effectively use the capital. If it exceeds the positive limit, positions on the long
stocks are increased. Similarly, if the TA exceeds the negative limit, long positions are
trimmed to bring back the account balance to zero. It would be preferable to shift
weight during rebalancing to the stocks with the most expected momentum left. How-
ever, since we only have +10 -10 stocks and a small sample to pick from, the portfolio
turns out to be very undiversified in these simulations.
• All trades are made at closing prices, and all changes in the portfolio are made within
the day. For example, if a long position exceeds the stop-loss limit, it will be sold at a
worse price than the set limit (the closing price), and a new stock will be added that
day at that day’s closing price.
• The portfolios are rebalanced (in terms of total weight of long and short portfolios)
every time the long or the short portfolio exceeds the other with more than 25%, in
order to keep market neutrality (see Tables 6 and 7). If the long portfolio outgrows
the short, the short portfolio is increased in value by adding to existing positions. The
same process is applied if the short portfolio outgrows the long, which happens rarely.
However, this assumption leads to a small asymmetry between the portfolios where the
short portfolio on average is worth less than the long. This is not the case in the (less
realistic) classical methodology with CAR where an equal money amount is assumed
when the portfolio is periodically rebalanced.
• Interest earned on the proceeds from the shorting of stocks (which is used as a security
by the broker) is set to 0%. This would in real life be somewhat less than the risk-free
interest rate. Interest rate on the TA is also set to zero.
• Fees on entering short positions are set to 0%. This would be less than 0.5% on
an annual basis for the more liquid stocks, but over 10% for illiquid stocks that are
drastically declining in price for longer periods (this happens rarely according to Germer
(2002)). Not including these fees, and the interest rate on the Transaction Account, is
however more than offset on average by not including the interest rates earned on the
security deposits described above.
• We start each simulation with a nominal value of 10 million SKr. The absolute starting
level is however irrelevant for the statistical analysis since everything is expressed in
relative terms and there are no fixed costs assumed.
6 RESULTS - PORTFOLIO SIMULATIONS 29
6 Results - Portfolio Simulations

6.1 Results of Our Trading Model
The idea now is to apply a battery of descriptive statistical tests to evaluate the performance
of our model. These are described in Table 13. We believe that all of these measures are
necessary complements to each other, although not all are discussed in this thesis. The
futility of using only CAR and a t-statistic when evaluating trading strategies is evident.
Unfortunately, it is very common.
7 Development of Our Default Model

x 10
4.5
AFGX
AFGX 3m Moving Average
4 Our Model 3m Moving Average
Our Model
Net Position Short
3.5
2.5
Skr
1.5
0.5
−0.5
01/01/95 07/02/97 01/01/00 07/02/02
Date
Figure 6. A diagram over the development of our model, compared with the same investment in
AFGX.
Table 13. Evaluation variables.
Evaluation Variable Mathematical Definition Comment on variable
Annualized Return (1st “ Total Value ”1/T How much money can this strategy produce
t=T ,T = num-
moment) Total Valuet=0 per year?
ber of years since inception.
Daily Return day t Total Value

t+1
R = log( Total Value ), t = day
t
√
Annualized standard de- (monthly standard deviation) 12 We have used months as a base for calculat-
viation (2nd central mo- ing annualized standard deviation since the
ment) trading horizon is a few months. Standard
deviation alone do not however give a cor-
rect picture of a portfolio’s risk characterist-
ics. Also, there are many ways to measure
standard deviation, which makes it suscept-
ible to manipulation. Had we used 3 months
averaged portfolio value as a basis for cal-
culating std, the Sharpe ratio would be over
5!
Skewness (3rd moment) See Appendix This measure and kurtosis is based on daily
values. Negative skewness is less attractive
than positive.
Kurtosis (4th moment) See Appendix The lower kurtosis, the better, since that will
mean less exposure to extreme events.
Sharpe (rf = 4%) Annualized Return−4% We have used 4% as the risk-free interest
Annualized Std rate. This measure is rather blunt for our
kind of strategy since it does not consider
correlations or distinguish between positive
or negative (return) contributions to volatil-
ity. However, as a comparison tool for sim-
ilar simulations it is excellent.
VaR 1 day (worst, 1%, 5%, The Value at Risk measure can be used to
10%) get a feeling for the negative end of the dis-
tribution of returns.
Worst 1 day, 1 week, 1 A good way to establish how bad bad can
month, 1 year) get.
“S ”
Worst/Best log T Date given is starting date for period. Note
St that the returns are logged, which e.g. means
that the continuous return one year can be
below -100% (or -63.2% =e−1 − 1 simple re-
turn).
Downside Risk Pt=T 2 This measure gives a value to the aggreg-

t=0 (min(0, daily returnt −rf )) ,
t = day ate returns below the risk-free rate (averaged
yearly risk-free rate in Sweden).
Beta (AFGX) See Appendix Comovement with market. A low β provides

insurance which is valuable to investors.
Autocorrelation total re- See Appendix The return’s correlation with itself yields an
turn (1 day, 1 week 1 indication of whether the portfolio is prone
month) to trend.
Accumulated transaction This measure quantifies total transaction

costs divided by final total costs. One should divide with time elapsed
portfolio value since inception and also consider the depend-
ence of ending total value to make this num-
ber comparable across simulations.
Total return from short This non-normalized measure before transac-

positions and total return tion costs (total sum of percentage return of
from long positions all positions) tells us the relative decomposi-
tion of returns from the long and short port-
folio. The short portfolio will lose money on
average and can be interpreted as the price
to be paid for lower risk and beta.
The results of our model is displayed in Table 14. In Figures 7 and 8 the distribution of
returns is depicted.
200
180
160
140
120
Frequency
100
80
60
40
20
0
−8 −6 −4 −2 0 2 4 6 8
Daily Return (%)
Figure 7. Histogram describing the distribution of the daily returns of our model.
15
10
Frequency
0
−20 −15 −10 −5 0 5 10 15 20 25
Monthly Return (%)
Figure 8. Histogram describing the distribution of the monthly returns of our model.
Table 14. Results of our trading model.

Sweden
Number of stocks 94-175
Time period 19950630-20011106
Annualized Return 21.8%
Daily Std 1.46%
Annualized Std 25.6%
Downside Risk 4.33%
Beta -0.011
Skewness -0.154
Kurtosis 2.20
VaR (daily)
Worst -6.90%
1% -3.95%
5% -2.26%
10% -1.57%
Autocorrelation
One day 0.0981
One week 0.0248
One month -0.0307
Sharpe (rf = 4%) 0.71
Acc. Tr. Costs/Total Value end 13.6%
Acc. Return Long 2181
Acc. Return Short -159
No. of short/long investments/run 344/272
Worst annual returns 1996-06-30: -7.15%
1998-06-30: -0.705%
1995-06-30: 9.44%
1999-06-30: 17.9%
2000-06-30: 41.4%
Worst monthly returns 1998-12-28: -19.2%
2001-09-28: -12.1%
1998-09-28: -11.8%
2000-04-28: -8.96%
1999-08-28: -8.44%
Best annual returns 1997-06-30: 64.4%
2000-06-30: 41.4%
1999-06-30: 17.9%
1995-06-30: 9.44%
1998-06-30: -0.705%
Table 15. Correlations of our trading model with different industries.

Industry Correlation with our model Correlation with our model from 980101
Bank and financial 0.01653 -0.02286
Forestry -0.03130 -0.03283
IT N/A 0.00353
Manufacturer -0.01962 0.01270
Miscellaneous 0.00669 0.03007
Retail 0.01143 0.03245
Chemicals and pharma -0.02407 -0.09324
Investment companies -0.01788 0.00585
Real estate and constr -0.00729 0.01093
Shipping and transportation -0.02769 -0.00090
6.1.1 Comments on the Results of Our Model

The stock market in Sweden between 1995 and 2001 is characterized by a relatively calm
period followed by a boom-bust cycle, which should be a favorable environment for mo-
mentum traders that manage to time the turnaround in the market. As one can see in the
Appendix, without risk control, our momentum strategy produces returns close to 40% per
year in this period. But it is not entirely straightforward to objectively compare the merits
of different strategies, and return is certainly not the only (or even most important) measure
of performance.
To establish the optimal balance between return, risk structure and correlations, one
would need to specify a utility function, which we have not done due to the inherent sub-
jectivity and difficulties in doing that. Instead, we have simply set out that we would like to
produce returns that are comparable to the naïve momentum strategies while avoiding large
losses and sustained periods with negative returns. But we have only roughly optimized the
model; we could for example have fine-tuned the sector ban rules to strike in the beginning
of March 2000 and hence produced better results.

The basic problem we have had with our strategy is that the model at times bans up
to 40% of an already small sample. We have chosen to accept this flaw and recognize that
results would be smoother and more diversified in a larger market with more stocks to choose
from. Also, as we stated earlier in Section 4, the fundamental data is of poor quality before
1997. Other than that, the result depicted in Table 14 is rather representative of our kind
of strategy.
Before 1998, the market moves sideways and the model has a hard time finding profitable
positions in stocks with momentum. Return picks up in 1998 until the Russian crisis shakes
the markets. Then comes a period of rather poor performance where there are no clear
trends in the market. Our risk control rules ban most of the stocks that skyrocketed in 1999
from entering the long portfolio (more than doubled M/B in 6 months). In the beginning of
2000, some of these technology stocks have entered the portfolio, only to start losing money
in March. And from the summer of 2000 to the last trading day, a lot of value is created
from short positions in dwindling stocks in the bust tech sector. In summary, we can see the
importance of clear trends and volatility in the underlying market for the profitability of our
strategy.
All things considered, our strategy has performed very well in this period, delivering both
high returns and a low correlation with the market index.
6.1.2 Comparison with the Simplistic Momentum Strategy

The simplistic strategy (see also Appendix A.6 for a definition) common in academic papers
on the momentum effect means basically a static rebalancing after a fixed interval and ranking
based purely on the price variable. In this strategy, we change all stocks every 4th month
and go long (short) in the 10 stocks that have exhibited the biggest increase (decrease) in
price the last 6 months. This is thus the strategy we set out to compare our own model
with as stated in the purpose of this thesis, and also the strategy that usually has the best
performance in academic research (see Section 2.4). We want to find out how an inclusion
of transaction costs and more advanced trading rules affect performance. (Our emulated
simplistic strategy still differs somewhat from the strategies in Section 2.4, due to our use of
the Transaction Account and rebalancing rules. This is however of minor importance.)
It is important to note that we have already assumed, based on theoretical and empirical
evidence, the existence of a momentum effect. Our goal here is merely an assessment of how
performance is affected when introducing the added realism of transaction costs and more
advanced trading rules in relation to a simplistic strategy. For a discussion of the strength
of this underlying assumption, see our analysis of robustness with respect to our sample in
Appendix A.1.
The robustness of the relationship between our own model and the simplistic strategy
is however another question. The inclusion of transaction costs always lowers returns while
keeping risk virtually unchanged. The use of more advanced trading rules always lowers
risk in approximately the same magnitude as evident here, and some return is of course also
sacrificed. This conclusion is supported both by theoretical arguments and statistical tests of
our simulations (see Appendix A.2). The robustness of this relationship is therefore stronger
than that of the model itself, which strengthens our conclusions.
Looking at Table 16 and comparing it with Table 14, we can see that even before in-
cluding transaction costs, the Sharpe measure is about the same (0.68 compared with 0.71).
Sharpe incorporates the two first moments of return and is hence a good indicator of relative
performance. The rest of the statistics tells the story why this is so and should not come
as a surprise. The risk characteristic of the simplistic strategy is dramatically worse than
in our model. All the risk measures; standard deviation, downside risk, VaR, are worse. In
Appendix A.2, we find that the standard deviation of monthly returns is indeed significantly
lower in our model. Returns on a monthly basis are not significantly different however, with
or without transaction costs, leaving Sharpe virtually unchanged. Although we do see some
tendencies of lower return in our model with transaction costs compared with the simplistic
model without transcation costs, these tendencies are not as statistically significant as the
improvement in risk.
The use of more advanced trading rules in our model have thus paid off, and the bulk
of improved performance is apparent on the risk side. We do give up some upside potential
when controlling risk, but this and the effect of transaction costs is compensated for by the
much better risk profile in terms of risk-return tradeoff. There are many things that can be
said about the relative performance, but these facts are the most important.
We can thus conclude that, when assessing the feasibility of a simplistic no transaction
cost momentum strategy commonly depicted in academic research in a real-life setting, per-
formance loss when introducing transaction costs may be less than the improvement from
more advanced trading rules. In our case and in this sample, assuming any risk-averse utility
function, the attractiveness of the risk-return profile has actually improved a lot due to the
lower risk.
Table 16. Results of simplistic momentum strategy.

Simplistic price mo- Simplistic price mo-
mentum without trans- mentum with transac-
action costs (replication costs
tion of earlier studies)
Number of stocks 101-175 101-175
Time period 19950602-20011106 19950602-20011106
Annualized Return 30.4% 19.8%
Daily Std 2.52% 2.47%
Annualized Std 39.1% 40.9%
Downside Risk 7.93% 8.91%
Beta -0.062 -0.076
Skewness -0.707 -2.07
Kurtosis 22.5 34.5
VaR (daily)
Worst -26.1% -29.7%
1% -7.30% -7.00%
5% -3.57% -3.29%
10% -2.07% -1.96%
Autocorrelation
One day 0.194 0.171
One week 0.012 -0.019
One month 0.147 0.157
Sharpe (rf = 4%) 0.68 0.39
Acc. Tr. Costs/Total Value end 0% 11.4%
Acc. Return Long 2887 2887
Acc. Return Short -274 -274
No. of short/long investments/run 200/200 200/200
Worst annual returns 1999-06-30: 8.68% 1999-06-30: -32.1%
1995-06-30: 13.1% 1995-06-30: 9.91%
1998-06-30: 16.9% 1998-06-30: 15.1%
1996-06-30: 18.7% 1996-06-30: 16.3%
2000-06-30: 29.7% 2000-06-30: 31.9%
Worst monthly returns 2000-12-28: -46.2% 2000-12-28: -45.9%
1999-12-28: -27.3% 1999-12-28: -44.3%
2001-09-28: -19.8% 2001-09-28: -21.6%
1999-02-28: -14.7% 2000-01-28: -13.8%
2001-04-28: -13.6% 2000-04-28: -12.9%
Best annual returns 1997-06-30: 72.8% 1997-06-30: 69.8%
2000-06-30: 29.7% 2000-06-30: 31.9%
1996-06-30: 18.7% 1996-06-30: 16.3%
1998-06-30: 16.9% 1998-06-30: 15.1%
1995-06-30: 13.1% 1995-06-30: 9.91%
7 CONCLUSIONS 35
7 Conclusions
There is both theoretical and empirical support for the existence of the momentum effect.
But any good quantifications of what performance actually can be achieved in a real trading
strategy are hard to get by. Results presented in most academic journals are not based on a
methodology realistic enough to measure the performance really available to investors.
We have found that, when introducing transaction costs and more advanced trading rules
using Swedish data from 1995 to 2001, performance as measured by e.g. Sharpe does not
deteriorate in relation to a simplistic strategy (with static rebalancing, no transaction costs,
only price as input variable and no other trading rules). The negative impact on return
when introducing transaction costs (0.3% per trade) is outweighed by the much lower risk
attainable with our model that uses a variety of data and rules to optimize performance.
In fact, the improved risk characteristics of our more realistic model infer that exploiting
the momentum effect might be even more attractive as a trading strategy than have been
suggested in academic research.
Having stated this conclusion, we would like to emphasize once again that the real im-
portance of this thesis lies in the use of the simulation platform, which took us 6 months to
develop. But it now takes us only 5 seconds to run a 10-year simulation, a few minutes to
adjust the parameters and test an alternative hypothesis, and a few hours to download new
data and test any other trading strategy. And all this is done in a way that is more realistic
than the use of for example CAR.
If the finance departments at business schools were to adopt an introduction to simulation
platforms in their methodology courses, maybe someday one would not have to look up the
impact of transaction costs in a footnote in the Appendix.
A APPENDICES 36
A Appendices
A.1 Robustness
A.1.1 Four Criteria of Robustness
We will here discuss the robustness of our model and the momentum effect with the help of
simulations from Germany, France, UK and Sweden.
We will use four conditions of robustness to evaluate our trading model. These conditions
have implications for the profitability in a real life setting, and whether the strategy benefits
investors in a way that is at least as good as in the historical simulations. These four
stipulations are as follows:
1. Liquid Markets
One of the theoretical explanations for the existence of the momentum effect is illiquid
markets, or limits to arbitrage with respect to the short side. We have assumed that all
the stocks in the sample can be shorted without delay, and that there always is someone
willing to lend us stocks. This is not a realistic assumption indeed, but we believe that these
constraints will not impede the results if the investor is big enough to have good relationships
with brokers, but small enough not to drain the market or affect prices.
For the simulations on the Swedish markets, where the lowest volume stocks in our
sample have a turnover of about 1 million SKr per day, it is evident that a portfolio bigger
than 10 million SKr would not allow immediate execution as assumed in this thesis, see
below for further discussion. In the foreign countries, where we have stocks based on market
capitalization (>4.9 billion SKr), the liquidity is better. For example, the daily turnover of
the 200 UK stocks we have in our sample is above 90 million SKr. But in France, there are a
few high capitalization, low volume stocks that one realistically could not trade in. We have
commented on this in the simulations below.
To test the robustness of our default model in Sweden (where we have daily data on
volume), we exclude the 10% stocks with the lowest volume, and do a simulation where
these stocks are included, but subjected to delayed execution and higher transaction costs.
2. Trading rules based on sound empirical and theoretical evidence

We have seen that it is a trivial task to construct very profitable strategies. Try using this
rule starting 1992: buy large cap Finnish companies that begin with the letter ”N”. This is
obviously an ex post constructed rule, that is not based on a theory why it should outperform
the market. We have to have some kind of reason to believe that a certain strategy will
continue to work, and we have found that support in the theoretical and empirical findings
of behavioral finance. And the best way of checking for the potential of any sample specific
bias is to compare the results with previous studies and to do out-of-sample tests of the
model.
We have done this on German, French and UK data, by replicating the methodology (only
using price as momentum signal) of previous empirical studies and comparing the results with
randomly generated signals. For these countries, we have almost 12 years of data, and the
time consistency of performance can also be studied. In Sweden, we have rather few stocks
(for a 20 stock momentum portfolio) between 1990 and 1995. When we run simulations using
rules that can ban stocks or industries, it can happen that the portfolio almost runs out of
stocks to consider including. In general, assuming that 5-10% of the market at any given
moment exhibit momentum characteristics, including more than that in a portfolio is likely
to produce bad results. Hence, these simulations should be interpreted with care.
A APPENDICES 37
But the biggest threat for an absolute return manager that believes that there are real
arbitrage profits to be made is however new entrants. Theory describes a gloomy scenario in
the case of too many investors chasing too little momentum: markets will be more volatile
and it will be much harder to catch the momentum in time. The Catch 22 is this; one want
to have good and solid research to build your strategy on, but one does not want anyone else
to know about it. We have not tested for this problem but can only note that momentum
strategies appear profitable up to date and that behavioral inefficiencies are likely to persist.
3. Correct assessment of risk, especially extreme events

Historical data is a poor judge when deciding upon the merits of a model’s protection to-
wards extreme events. The negative tails are stochastic and hard to estimate correctly. We
believe that our default model is reasonably crash-proof if unleveraged, at least on the cata-
strophic scale. Changes in the volatility structure of the market may lead to periods of worse
performance, but not sudden crashes. But we can never be certain. For the naïve strategies
without risk control, big drops in value are common however. One can look at the VaR table
to get a feeling for the bad side of the return distribution. This requires that one has long
data series.
4. Correct assessments of costs incurred

We have included transaction costs, and made in our view assumptions on fees and interest
rate charged and earned that lower returns in our simulations. However, costs like getting
information, setting up computers and office space, taxes, and so forth are not included. In
one of the simulations below, we check for the sensitivity of the default model when increasing
transaction costs.
A.1.2 Tables of Results for Robustness

Since we only have the price variable for the German, UK, and French markets, we primarily
use this data to check for the momentum effect with roughly the same methodology as
indicated in Section 2.4. We just turn off all risk control, and let the model rank the stocks
according to price change the last 6 months. The holding period is set to 4 months. This
means that all stocks will be replaced simultaneously every 4th month.
The results presented in Table 17 are clearly consistent with previous studies, e.g. the
study on the German market by D. Schiereck and Weber (1999). An abnormal return of a
little bit less than 1% per month, together with a risky volatility structure with large and
sudden drops in value.
Without risk control, this naïve strategy is not very interesting for an investor, even
before considering transaction costs. But for academic purposes, the results strengthen the
hypothesis that there is return continuation in the medium term.
There are two problems with the data presented here, one is the few stocks in the begin-
ning of the Swedish sample, and one is the existence of thinly traded stocks in the French
sample. One exceptional example of a winning position is noticeable on the first of August
2000 when a stock called Look Voyages increases from 3 to over 40 franc in one day (which
helps explain the large skewness for French stocks). The company is a 97% majority owned
travel group. This experience is also a hint for a rule that bans short positions in stocks that
are not widely held. If we exclude the four French stocks that have increased or decreased
more than 50% in one day, the annual return stops at 2.55% and annual standard deviation is
33.2%. In Sweden, where the devaluation of the currency shook up the economy in 1992-93,
this momentum strategy lost over 70% (simple return) in one year (see Section 2.4).
A APPENDICES 38
Due to these problems, we would like to consider the UK and German simulations to
have the most reliable samples.
Table 17. Replication of earlier studies. Settings: No transaction costs, no risk control, ranking
based on last 6 month price change (not weighted), rebalancing every fourth month.
Germany UK France Sweden
Number of stocks 110-167 192- 202 110-181 90-175
Time period 19900702-20011106 19900702-20011106 19900702-20011106 19900702-20011106
Annualized Return 8.25% 10.9% 7.42% 13.9%
Daily Std 1.22% 1.32% 2.24% 2.57%
Annualized Std 22.5% 27.7% 42.7% 44.9%
Downside Risk 10.6%
Beta -0.095
Skewness -0.266 -1.19 18 -0.346
Kurtosis 4.36 9.94 6.89 22.66
VaR (daily)
Worst -15.5% -11.4% -24.0% -32.9%
1% -3.36% -4.20% -4.88% -7.39%
5% -1.82% -2.09% -2.19% -3.16%
10% -1.21% -1.39% -1.56% -2.09%
Autocorrelation
One day 0.0427 0.2040 0.0168 0.0875
One week 0.0118 0.0394 0.0166 -0.0155
One month 0.0139 0.0072 0.0013 0.0116
Acc. Tr. Costs/Total Value end 0% 0% 0% 0%
Sharpe (rf =4%) 0.19 0.25 0.08 0.22
No. of short/long investments 350/350 350/350 350/350 350/350
Acc. Return Long 1854 2717 3104 2887
Acc. Return Short -716 -1057 -1505 -274
Worst annual returns 1992-07-02: -24.3% 1998-07-02: -38.3% 1998-07-02: -53.7% 1992-07-02: -161%
1998-07-02: -6.95% 1999-07-02: -31.3% 1992-07-02: -46.5% 1999-07-02: -37.9%
1991-07-02: -2.09% 1992-07-02: -13.9% 1990-07-02: -42.0% 1990-07-02: 8.19%
1990-07-02: -1.83% 1993-07-02: -1.07% 1993-07-02: -12.4% 1995-07-02: 10.4%
1997-07-02: 3.11% 1990-07-02: 5.09% 1997-07-02: 9.84% 1994-07-02: 14.4%
Worst monthly returns 2001-04-02: -19.1% 2000-05-02: -33.2% 2000-03-02: -30.5% 1992-11-02: -65.1%
2000-03-02: -16.7% 2000-03-02: -32.2% 2001-10-02: -25.1% 1993-02-02: -62.1%
2001-10-02: -15.9% 1999-04-02: -24.3% 2001-04-02: -24.9% 2000-03-02: -36.5%
1998-05-02: -12.9% 2001-10-02: -17.3% 2000-05-02: -23.8% 2001-10-02: -31.5%
1997-05-02: -12.9% 1991-02-02: -17.2% 1993-02-02: -23.3% 2000-05-02: -31.4%
Best annual returns 1995-07-02: 37.2% 1991-07-02: 53.1% 2000-07-02: 112% 1997-07-02: 99.5%
2000-07-02: 28.8% 2000-07-02: 46.0% 1999-07-02: 39.2% 2000-07-02: 68.8%
1999-07-02: 21.4% 1995-07-02: 37.9% 1995-07-02: 32.5% 1991-07-02: 68.4%
1994-07-02: 14.7% 1997-07-02: 23.7% 1996-07-02: 31.6% 1998-07-02: 38.2%
1993-07-02: 14.5% 1996-07-02: 17.2% 1994-07-02: 22.8% 1996-07-02: 19.5%
The significance of the momentum effect is clearly evident when comparing Table 17 with
Table 18, where we have ranked the stocks randomly. The average return is centered slightly
below zero due to the asymmetry between the long and short positions (see Section 5.3.1).
Returns are not as volatile here because of the lower correlations within the portfolios.
A APPENDICES 39
Table 18. Random rankings (median of 50 runs). Settings: No transaction costs, no risk
control, ranking based on a random number generator, rebalancing every fourth month.
Number of stocks 110-167 192-202 110- 181 90-175
Time period 19900702-20011106 19900702-20011106 19900702-20011106 19900702-20011106
Annualized Return -0.75% -1.81% -1.31% -0.65%
Daily Std 0.739% 0.813% 0.981% 1.32%
Annualized Std 11.5% 13.4% 14.4% 21.0%
Downside Risk 4.72%
Beta -0.008
Skewness -0.141 0.522 -0.796 1.02
Kurtosis 2.62 8.798 9.76 20.572
VaR (daily)
Worst -5.99% -5.05% -10.1% -7.90%
1% -1.86% -2.18% -2.67% -3.53%
5% -1.14% -1.21% -1.47% -2.08%
10% -0.86% -0.913% -1.06% -1.51%
Autocorrelation
One day 0.0249 0.0941 0.0080 0.0380
One week 0.0085 -0.0384 -0.0221 -0.0158
One month 0.0211 -0.0038 -0.0183 0.00822
Sharpe (rf = 4%) -0.41 -0.43 -0.37 -0.22
Acc. Tr. Costs/Total Value end 0% 0% 0% 0%
Acc. Return Long 829 1845 871 1014
Acc. Return Short -883 -1732 -1269 -1089
No. of short/long investments/run 350/350 350/350 350/350 350/350
Worst annual returns 1997-07-02: -24.4% 1991-07-02: -32.7% 1992-07-02: -16.5% 1999-07-02: -58.4%
1995-07-02: -16.5% 1995-07-02: -13.3% 1990-07-02: -14.0% 1991-07-02: -30.9%
1999-07-02: -2.62% 1996-07-02: -6.57% 1998-07-02: -11.1% 1994-07-02: -13.4%
1996-07-02: 1.87% 1992-07-02: -5.10% 2000-07-02: -10.8% 1998-07-02: -3.19%
1994-07-02: 2.39% 1990-07-02: -5.09% 1999-07-02: -9.36% 2000-07-02: 1.17%
2000-01-02: -10.1% 1995-09-02: -11.8% 2000-01-02: -15.1% 1999-11-02: -17.7%
2001-10-02: -7.75% 1995-11-02: -9.28% 1999-12-02: -14.7% 1999-12-02: -14.2%
1993-08-02: -7.64% 1992-02-02: -8.84% 1997-02-02: -9.05% 1999-10-02: -12.1%
1998-02-02: -7.41% 1991-09-02: -8.46% 1990-10-02: -8.44% 1992-04-02: -11.8%
Best annual returns 1992-07-02: 14.8% 1998-07-02: 24.2% 1996-07-02: 10.5% 1993-07-02: 27.8%
2000-07-02: 11.1% 1999-07-02: 9.11% 1994-07-02: 7.33% 1992-07-02: 25.9%
1990-07-02: 8.99% 1993-07-02: 8.70% 1997-07-02: 3.17% 1997-07-02: 22.2%
1991-07-02: 6.70% 2000-07-02: 7.43% 1991-07-02: 0.21% 1990-07-02: 16.1%
1993-07-02: 3.65% 1994-07-02: 5.42% 1993-07-02: -0.24% 1996-07-02: 12.7%
In Table 19, we have introduced transaction costs and risk control with the price variable.
The portfolio will now quickly move out of phase because of the stop-loss rule, and we can
see that the turnover is higher. The returns still seem to be significantly different from zero
after transaction costs have been subtracted. But the volatility has not come down much at
all. Again, France has important outlier observations that affect the results, when we delete
the 4 worst stocks, annualized return goes down to 6.05% and annual standard deviation
to 26.9%. This makes the results across the European markets we have tested look very
similar. The higher volatility in Sweden can be explained by the much smaller and more
volatile stocks in the Swedish sample.
A APPENDICES 40
Table 19. Transaction costs and risk control based on price variable. Settings: 0.3%
transaction costs, risk control (stop-loss 25% of maximum price, covariance matrix), ranking based
on weighted return of 6 months price changes (last week: 20%, rest: 80%), stocks are held until
stop-loss or 4 months pass.
Number of stocks 110-167 192-202 110-181 90-175
Time period 19900702-20011106 19900702-20011106 19900702-20011106 19900702-20011106
Annualized Return 5.77% 7.75% 15.1% 11.2%
Daily Std 1.27% 1.25% 2.54% 2.00%
Annualized Std 24.5% 25.4% 43.9% 35.9%
Downside Risk 7.63%
Beta -0.024
Skewness -0.666 -0.0222 27.3 -0.613
Kurtosis 7.46 10.032 118 6.07
VaR (daily)
Worst -8.53% -10.2% -9.55% -37.0%
1% -3.86% -3.92% -4.22% -5.69%
5% -1.95% -1.93% -2.05% -2.65%
10% -1.38% -1.30% -1.44% -1.83%
Autocorrelation
One day 0.119 0.165 0.0439 0.119
One week -0.0189 0.031 0.00822 0.0304
One month -0.0309 0.003 -0.0048 0.0226
Sharpe (rf = 4%) 0.07 0.15 0.25 0.20
Acc. Tr. Costs/Total Value end 22.3% 36.9% 11.3% 14.4%
Acc. Return Long 2029 2897 3731 2264
Acc. Return Short -980 -1402 -1329 -471
No. of short/long investments/run 429/380 495/376 481/391 601/420
Worst annual returns 1990-07-02: -30.3% 1999-07-02: -25.1% 1990-07-02: -43.3% 1992-07-02: -33.1%
1998-07-02: -15.2% 1993-07-02: -0.80% 1998-07-02: -38.9% 1990-07-02: -29.1%
1996-07-02: -7.35% 1994-07-02: -0.67% 1992-07-02: -14.7% 1993-07-02: -23.1%
1997-07-02: -6.51% 1997-07-02: -0.05% 1996-07-02: -4.37% 1998-07-02: -22.2%
1992-07-02: -5.18% 1996-07-02: 0.460% 1994-07-02: 10.9% 1995-07-02: 2.35%
2000-03-02: -25.8% 2000-03-02: -22.0% 2001-10-02: -23.3% 1991-01-02: -34.1%
1991-02-02: -16.6% 1997-08-02: -15.6% 2000-08-02: -20.8% 2000-03-02: -26.9%
1999-04-02: -13.5% 2000-04-02: -14.1% 2000-05-02: -15.8% 1999-09-02: -20.6%
2000-05-02: -12.4% 2000-05-02: -13.7% 1999-02-02: -14.6% 2000-08-02: -20.5%
Best annual returns 1995-07-02: 47.9% 1992-07-02: 41.1% 2000-07-02: 139% 1997-07-02: 68.5%
2000-07-02: 38.7% 1991-07-02: 37.9% 1995-07-02: 34.1% 1999-07-02: 44.3%
1993-07-02: 27.1% 1995-07-02: 32.5% 1991-07-02: 22.5% 2000-07-02: 41.6%
1994-07-02: 8.16% 2000-07-02: 16.6% 1993-07-02: 21.8% 1991-07-02: 25.5%
1991-07-02: 6.23% 1990-07-02: 4.41% 1999-07-02: 20.6% 1994-07-02: 21.4%
Looking only at the period used in our default model, we can see that the relationships
between the different types of simulations persist. However, returns are about twice as high
in this period, as evidenced by the data in Table 20.
A APPENDICES 41
Table 20. Simulations from 1995 in Sweden.

Replication of earlier Random ranking Transaction costs and
studies (Table 17) (Table 18) risk control with price
variable (Table 19)
Number of stocks 101-175 101-175 101-175
Time period 19950602-20011106 19950602-20011106 19950602-20011106
Annualized Return 30.4% -3.28% 24.0%
Daily Std 2.52% 1.23% 1.84%
Annualized Std 39.1% 17.3% 33.9%
Downside Risk 7.93% 3.18% 5.99%
Beta -0.062 0.052 0.0064
Skewness -0.707 1.02 -0.613
Kurtosis 22.5 20.6 6.07
VaR (daily)
Worst -26.1% -7.53% -10.6%
1% -7.30% -3.02% -5.98%
5% -3.57% -1.89% -2.78%
10% -2.07% -1.38% -1.82%
Autocorrelation
One day 0.194 -0.0113 0.158
One week 0.012 -0.0250 -0.005
One month 0.147 -0.0286 0.044
Sharpe (rf = 4%) 0.68 -0.42 0.59
Acc. Tr. Costs/Total Value end 0% 0% 11.2%
Acc. Return Long 2887 1014 2264
Acc. Return Short -274 -1089 -471
No. of short/long investments/run 200/200 200/200 200/200
Worst annual returns 1999-06-30: 8.68% 1998-06-30: -39.5% 1998-06-30: -29.5%
1995-06-30: 13.1% 1995-06-30: -20.1% 1995-06-30: 0.195%
1998-06-30: 16.9% 1997-06-30: -19.4% 1996-06-30: 20.1%
1996-06-30: 18.7% 2000-06-30: 2.11% 1999-06-30: 41.1%
2000-06-30: 29.7% 1996-06-30: 17.7% 2000-06-30: 44.2%
Worst monthly returns 2000-12-28: -46.2% 1998-11-28: -9.55% 1999-08-28: -21.9%
1999-12-28: -27.3% 1998-07-28: -9.28% 1999-04-28: -16.5%
2001-09-28: -19.8% 1999-02-28: -8.97% 1998-12-28: -16.3%
1999-02-28: -14.7% 1998-04-28: -8.69% 2000-04-28: -16.0%
2001-04-28: -13.6% 1995-10-30: -8.32% 2001-09-28: -15.0%
Best annual returns 1997-06-30: 72.8% 1999-06-30: 33.3% 1997-06-30: 54.5%
2000-06-30: 29.7% 1996-06-30: 17.7% 2000-06-30: 44.2%
1996-06-30: 18.7% 2000-06-30: 2.11% 1999-06-30: 41.1%
1998-06-30: 16.9% 1997-06-30: -19.4% 1996-06-30: 20.1%
1995-06-30: 13.1% 1995-06-30: -20.1% 1995-06-30: 0.195%
Finally, we look at alterations of our default model in Table 21. The results are expected,
at least the direction of changes. No risk control yields a (much) higher return but also
higher risk. We do not interpret the higher Sharpe ratio as this being a better model. Apart
from worse risk characteristics, this profile seem to be rather time specific, considering the
extremely profitable trades in connection with the rise and fall of the technology sector. Also,
our method of banning stocks as a means for risk control in our default model is, as stated
above, not very optimal in a small sample.
Transaction costs (TC) of 0.7% significantly lowers the return. Note that the model will
not do the exact same trades as in the default model, when TC are altered. Much higher TC
than 0.7% would in any way be critical to the profitability of the model. When we introduce
delayed execution and higher TC for the 15 most illiquid stocks results are worse, but not
much. In our simulations, the buy/sell order is postponed 1 day for these stocks. The effect
of higher TC is thus more important in this case.
If we are restricted from trading in stocks with low volume, say the bottom 10% (less than
about 1.5 million SKr/day), return drops. But this simulation points more to the importance
of a large sample to start with rather than the importance of high volume, since there is not
a big difference in daily turnover of the stocks at the bottom of the chart. If one takes away
10%, and these being mostly growth companies, from a small sample, performance will be
severely hit.
An interesting simulation is if we just turn off the stop-loss rule (not tabulated). The
model will rebalance all stocks simultaneously every 4th month and thus not trade in the
same stocks as in the default model. We find that annual return is up to 30.9% and that the
VaR and standard deviation (30.3%) are higher. Most of our simulations exhibit improved
A APPENDICES 42
performance with the stop-loss rule. But these gains are most often on the risk side and are
most significant in the presence of long sustained trends that diminish portfolio value.
In a simulation where we only have used weighted price as a signal generator (not tab-
ulated), ceteris paribus, the Sharpe ratio is down to 0.04. We have found that when the
risk control rules are turned off, price is the most important signal generator, but if it is not
turned off, it does not work very well alone.
If we let the default model start the simulation in 1997 (not tabulated), the portfolio is
in essence the same as in a simulation where one starts in 1995 and cuts off the first two
years. In this case, performance is somewhat better because of the flat performance between
1995 and 1997 in our default model.
Also, the model does remarkably well in 1992-93, where the simplistic momentum strategy
lost over 70% simple return (see Table 17), even without access to fundamental data. Total
value is down 2.5% in the corresponding year.
Table 21. Changes in our default model. Default settings: 0.3% transaction costs, all ranking
signals and risk control rules
No risk Control 0.7% transaction costs 1 day delayed trades No trades in the 10%
and 0.6% higher trans- with lowest volume
action costs for 10% (<about 2 million
most illiquid stocks SKr/day)
Number of stocks 94-175 94-175 94-175 94-175
Time period 19950630-20011106 19950630-20011106 19950630-20011106 19950630-20011106
Annualized Re- 39.8% 13.4% 12.9% 17.0%
turn
Daily Std 1.94% 1.55% 1.42% 1.60%
Annualized Std 33.6% 31.5% 28.1% 31.6%
Downside Risk 5.82% 5.13% 5.56% 6.18%
Beta 0.0022 -0.014 -0.022 -0.053
Skewness -0.524 -0.416 -0.279 -0.373
Kurtosis 5.95 4.79 5.97 5.52
VaR (daily)
Worst -14.9% -10.9% -11.5% -11.8%
1% -5.87% -4.50% -4.24% -4.52%
5% -2.79% -2.26% -2.11% -2.39%
10% -1.87% -1.65% -1.49% -1.62%
Autocorrelation
One day 0.0830 0.137 0.128 0.126
One week 0.0462 0.0313 0.0659 0.0481
One month 0.0423 -0.0271 0.00822 0.0177
Sharpe (rf = 1.06 0.298 0.319 0.411
4%)
Acc. Tr. 7.46% 49.7% 20.5% 15.9%
Costs/Total
Value end
Acc. Return 2868 1952 2044 2436
Long
Acc. Return 173 -314 -436 -505
Short
No. of 200/202 342/268 343/265 339/270
short/long in-
vestments/run
Worst annual 1998-06-30: 0.860% 1998-06-30: -19.1% 1999-06-30: -21.4% 1998-06-30: -25.4%
returns
1996-06-30: 9.78% 1996-06-30: -12.2% 1996-06-30: -6.99% 1996-06-30: 2.60%
1995-06-30: 25.8% 1995-06-30: 3.85% 1998-06-30: 11.8% 1995-06-30: 7.93%
1999-06-30: 36.9% 1999-06-30: 18.3% 1995-06-30: 17.7% 1999-06-30: 25.8%
2000-06-30: 69.3% 2000-06-30: 33.2% 2000-06-30: 29.4% 2000-06-30: 26.7%
Worst monthly 2000-12-28 -28.1% 2000-02-28: -17.6% 2000-12-28: -23.1% 2000-12-28: -32.4%
returns
2001-09-28: -21.5% 1998-09-28: -15.9% 2000-02-28: -15.9% 2000-04-28: -17.1%
2000-02-28: -19.6% 2000-03-28: -12.6% 2000-03-28: -14.4% 1999-01-28: -16.8%
1999-01-28: -12.5% 2001-09-28: -12.6% 2000-04-28: -13.7% 2000-08-28: -13.6%
2000-03-28: -9.67% 2000-07-28: -11.4% 2000-08-28: -13.2% 1998-12-28: -12.9%
Best annual re- 1997-06-30: 77.2% 1997-06-30: 59.8% 1997-06-30: 47.7% 1997-06-30: 62.2%
turns
2000-06-30: 69.3% 2000-06-30: 33.2% 2000-06-30: 29.4% 2000-06-30: 26.7%
1999-06-30: 36.9% 1999-06-30: 18.3% 1995-06-30: 17.7% 1999-06-30: 25.8%
1995-06-30: 25.8% 1995-06-30: 3.85% 1998-06-30: 11.8% 1995-06-30: 7.93%
1996-06-30: 9.78% 1996-06-30: -12.2% 1996-06-30: -6.99% 1996-06-30: 2.60%
A APPENDICES 43
A.1.3 Discussion on the Robustness of Our Model

To reach a conclusive statement about robustness and other statistical properties of a trading
model one needs extensive data from many countries and time series that are measured in
decades. We have neither. But for the simplistic momentum strategy using only price as a
variable, we can compare with earlier research and thus come to a more distinct conclusion.
Our results confirm other researchers’ findings on the momentum effect. Results are
similar for Sweden, UK, Germany and France and are consistent with both prior empirical
evidence and contemporary theory. A market neutral momentum strategy in its most simple
form can earn money. It is not a random walk. This conclusion holds true even after including
transaction costs. But the performance is very volatile with long periods of good performance
and sudden losses. This is why normal statistical methods to establish significance of results,
such as the t-statistic, is afflicted with problems. Especially if the time period is too short
and important extreme events are not captured in the sample.
One dimension of our analysis of robustness, the existence and similarity of the mo-
mentum effect across countries can thus be given an affirmative answer with respect to our
sample.
As for our own self-made model, we can not draw any strong conclusions about the
exact level of performance. Just over 6 years of data is simply too short, and there are no
comparable studies. Considering that our model is a combination of many strategies, it is
not either very likely that someone ever will test all possible permutations of these strategies.
We can only tentatively state our conclusions and not accurately quantify future statistical
properties of our strategy.
But we can be reasonably sure that our result is not a statistical fluke, since we have
based our trading rules on established theoretical and empirical research, and not fine-tuned
the model to fit a specific time-period. In conclusion, we find that the following things can
be said about the robustness of results of our default model:
• On average, measured over a long time period, we believe that our model can produce
economically significant results in the future. This conclusion is based on evidence
from earlier research on the momentum effect and the relative stability of our model.
However, the Swedish sample using only the price variable yields higher returns in the
period where we have simulated our default model, than from 1990. It may be the case
that one should expect an average return that relates in the same way to the results
from 1995, i.e. only about half of the returns or 10% per year for the default model.
The years around the turn of the millenium seem to have been favorable indeed for
a momentum trader; first one long upward trend among especially high-tech growth
companies, and then an equally persistent bearish trend. The fundamental determinant
and limiting factor of profitability for a momentum strategy is the existence of persistent
trends in the market. It is not hard to pick up on these trends in general once they
show up, but it is a delicate task to spot the turnarounds that quickly diminish the
value of the portfolio. How would our model have acted in earlier periods and, more
importantly, how will it behave in the future? One can not exactly know the answer
to that question.
• To implement our model on the Swedish market, one can not be too big. With a daily
turnover of around 1-2 million SKr for the stocks with the lowest volume, we can with
a rough estimation say that returns will be hit if total portfolio value exceeds about
30 million SKr. When we excluded the 10% lowest volume stocks, which would not
be immediately executed in a 30 million SKr portfolio without price impact, returns
dropped significantly. It is therefore imperative to diversify across many markets, not
A APPENDICES 44
only to reduce risk, but also to gain access to more liquid stocks. Then it is possible
to run a much larger portfolio. The size of the market one trades in is thus a limiting
factor for performance.
• Transaction costs are critical determinants of profitability due to the high turnover of
the portfolio. One needs to be big enough to avoid paying a fixed fee, and even if you
can avoid this, average transaction costs most probably will exceed 0.3%. This limit,
to reach 0.3%, is about 1 million SKr with an Internet broker today. Hence, we have
both an upper and a lower limit to consider for the size of the portfolio. But these
limits are obviously not critical for the interested momentum trader, who would seek
out much larger and more liquid stock markets than the Swedish one.
A APPENDICES 45
A.2 Statistical Comparison of Our Model and the Simplistic Model

In this section we will compare the risk and return parameters of our trading model to the
ones of the simplistic price momentum model including and excluding transaction costs,
respectively. More specifically, we will try to test if our model performs worse than the other
two strategies in two dimensions: standard deviation and return. Both tests are based on
monthly figures, primarily to ensure large sample sizes and thereby more powerful tests. We
have used monthly figures as they yield a large enough sample, and are closer to the trading
horizon than figures on a daily basis. One could also test the equality of the Sharpe ratios,
but the available tests have been criticized for their low power, see e.g. Jobson and Korkie
(1981).
A.2.1 Volatility
At first sight, one would handle this topic by applying the standard procedure of testing the
difference between two variances (the so called variance ratio test). According to Newbold
(1995), "[t]hat test is valid when independent random samples from the two populations
are available". However, in our case we are dealing with matched pairs (since we have run
the simulations during the same time period during which the same set of stocks has been
considered). Henceforth the parameter of our model will be denoted X, and the one of the
model we compare to Y . The test procedure consists of a test for Zero Population Correlation
(using Newbold’s terminology), where one uses the fact that the correlation between (X −Y )
and (X + Y ) is zero if and only if σx2 and σy2 are equal. Our onesided test then looks like
this:
H0 : σx2 = σy2 (11)

H1 : σx2 < σy2 (12)
Decision rule:
r
Reject H0 if p < −tn−2,α , (13)
(1 − r2 )(n − 2)
where r is the sample correlation, σx2 is the variance of our model and σy2 is the volatility of
the simplistic model without transaction costs. n is the sample size (76 in all tests) and α is
the significance level. tn−2,α is the number for which
P (tn−2 > tn−2,α ) = α, (14)
where the random variable tn−2 follows a Student’s t distribution with (n − 2) degrees
of freedom. Replace Y with Z to get the notation for the model with transaction costs.
Comparing our model to the simplistic model without transaction costs, we find that the
correlation is −0.454 and the corresponding test statistic is −4.39. The null hypothesis can
thus (see Table 22) be rejected already at the five percent level, in favour of the hypothesis
that our model has a lower variance, and thereby a lower volatility.
The correlation in the second test case is −0.466 yielding a test statistic of −4.53. The
null hypothesis can once again be rejected at the five percent level, and the conclusions above
apply here as well.
A.2.2 Return
Testing for the difference in returns is a simple matter of testing the differences between two
means. In line with the standard deviation tests, we will use the matched pairs methodology
A APPENDICES 46
presented in Newbold (1995). The test is set up as follows:
H0 : µx = µy (15)
H1 : µx < µy (16)
Decision rule:
d¯
Reject H0 if √ < −tn−1,α , (17)
sd / n
where µx and µy are the average monthly returns of our model and the simplistic model
without transaction costs, respectively. d¯ is the sample mean of the difference between X
and Y (here X and Y are the monthly return vectors). sd is the sample standard deviation.
Replace Y with Z to get the notation for the model with transaction costs.
The test statistic in the case where we compare the return of our model to the return of
the simplistic model where the transaction costs are excluded is −0.398 which implies that
the null hypothesis cannot be rejected (see Table 22). The corresponding statistic for the
comparison between the returns of our model and the simplistic model excluding transaction
costs is 0.155, and yet again we cannot reject the null hypothesis that the population means
are equal. Thus, statistically speaking, one cannot say that performance in terms of risk and
return deteriorates when one introduces transaction costs and the set of more advanced rules
we have discussed in this thesis.
Table 22. A brief summary of the hypotheses tested above. x corresponds to our model, y to the
simplistic model without transaction costs and z to the one with transaction costs.
H0 H1 Test Statistic Critical Value (5 %) Result
2 2 2 2
σx = σy σx < σy −4.39 −2.58 Reject H0
2 2 2 2
σx = σz σx < σz −4.53 −2.58 Reject H0
µx = µy µx < µy −0.398 −2.66 Cannot reject H0
µx = µz µx < µz 0.155 −2.66 Cannot reject H0
A APPENDICES 47
A.3 Screenshot of the Program
Figure 9. The program in action.

A APPENDICES 48
A.4 A Day of Trading

***New Trading Day***
1996-02-22 340
Transaction Account: 198889.75 (1.9625976 % of long portfolio).
The portfolio consists of the following stocks:
DORO (short) 1996-02-06.
SSABA (short) 1996-01-24.
ELEKTAB (long) 1995-11-23.
SANDVIK (short) 1996-02-02.
SKANDIA (long) 1995-10-30.
AKZONOBELSDB (short) 1995-11-14.
SKFB (short) 1996-02-02.
BERGMANBEVINGB (long) 1995-10-30.
HENNESMAURITZB (short) 1995-10-30.
TRELLEBORGB (short) 1995-10-30.
MUNKSJO (short) 1996-01-12.
ROTTNEROS (short) 1996-02-06.
SAS (long) 1996-01-30.
NOBELBIOCARE (long) 1996-02-19.
CUSTOSA (long) 1995-10-30.
NCCB (long) 1995-10-30.
ALLGONB (short) 1995-10-30.
FRONTECB (long) 1996-01-23.
OM (long) 1995-10-30.
IBSB (long) 1995-10-30.
Size of long: 1.0562329E7.
Size of short: 9329879.0
Time to divest HENNESMAURITZB (short).
Transaction Account before drop: 198889.75
Transaction Account after drop: 167891.78
Banned short stock HENNESMAURITZB
HENNESMAURITZB belongs to the sector: [retail$]
Divested the short stock HENNESMAURITZB at a price of 949319.06 SEK.
The Transaction Account is now 167891.78 SEK.
SCAB: Number of days between 1996-02-06 and 1996-02-22 is 16
SCAB is unbanned.
CONCORDIAMARITIMEB: Number of days between 1996-02-09 and 1996-02-22 is 13
HENNESMAURITZB: Number of days between 1996-02-22 and 1996-02-22 is 0
The following stocks are banned from the short portfolio:
CONCORDIAMARITIMEB
HENNESMAURITZB
Original candidate: NOKIASDBSEE
Considering the following stocks:
NOKIASDBSEE <=> 1.9383791E-5
KLIPPANS <=> 2.1233047E-5
TV4A <=> 2.2088336E-5
PRIF <=> 2.3530345E-5
At 1996-02-22 NOKIASDBSEE was shorted for 930188.9 SEK.
Size of long portfolio at end of trading day: 1.0562329E7
Size of short portfolio at end of trading day: 9310749.0
Total value decomposition:
lv (1.0562329E7) + nps (-339003.47) + transaction account (165092.81) = 1.0388419E7.
Checking development of market-to-book (1996-02-23).
Market-to-book for ACTIVEBIOTECHB has developed from 2.54 to 0.51 implying a ratio of 0.2007874.
ACTIVEBIOTECHB has therefore been banned from the short list.
Banned short stock ACTIVEBIOTECHB
A APPENDICES 49
A.5 A List of the Swedish Stocks Used
Table 23. Swedish stocks. The numbers in parentheses indicate the number of times the stock has
entered the long and short portfolio in our default model, respectively.
ABBSEE (0, 0) GORTHONLINES (1, 2) PANDOX (2, 1)
ACTIVEBIOTECHB (3, 3) GOTLANDREDERIB (4, 0) PEABB (4, 3)
AGA A (1, 4) GRANINGE (1, 0) PERS B (2, 4)
AGA B (3, 3) GUNNEBO (3, 0) PHARMACIASDB (0, 3)
AKZONOBELSDB (1, 1) HAGSTROMERQVIBERG (0, 0) PIRE (3, 2)
ALFA (2, 4) HALDEX (1, 5) PLAT B (1, 2)
ALLGONB (0, 5) HANDBKNHYPOTEKPREF (2, 0) PLM (1, 1)
ALTH B (2, 4) HENNESMAURITZB (0, 3) PREVASB (1, 3)
ANGPANNEFORENINGENB (6, 1) HEXAGONB (3, 1) PRICERB (1, 2)
ASSAABLOYB (3, 1) HLDISPLAYB (4, 2) PRIF (1, 4)
ASSIDOMAN (1, 0) HOGANASB (2, 2) PROTECTDATA (3, 2)
ASTI (2, 0) HOISTINTLB (4, 1) RATOSB (2, 0)
ASTRAZENECASEE (0, 1) HUFVUDSTADENA (0, 4) REALIAB (1, 2)
ATLASCOPCOA (0, 0) IBSB (1, 5) RESCOB (1, 2)
AUTOLIVSDB (0, 2) ICONMEDIALABINTL (2, 4) ROTTNEROS (1, 3)
AVESTAPOLARITSEE (1, 0) INDLFINLSYSB (2, 1) SAABB (0, 0)
BERGMANBEVINGB (4, 3) INDUSTRIVARDENA (0, 1) SALUSANSVARB (0, 1)
BIACOREINTL (3, 3) INTENTIAINTLB (1, 5) SANDVIK (1, 3)
BILIAA (1, 3) INVESTORB (0, 0) SAPA (0, 0)
BIOGAIABIOLOGICSB (2, 0) INVIKB (2, 1) SARDUS (0, 0)
BIOPHAUSIAA (3, 1) ITABINDUSTRIB (2, 0) SAS (2, 4)
BIORA (2, 4) JM (4, 0) SCAB (0, 1)
BNNORDSJB (0, 4) JPB B (0, 1) SCANIAB (2, 4)
BONGSLJUNGDAHLB (5, 0) KARLSHAMNS (1, 0) SCRIBONAB (1, 3)
BPA B (3, 1) KAROBIOB (1, 0) SEBA (0, 4)
BROSTROM (2, 2) KAROLINMACHINETOOL (0, 1) SECOTOOLSB (2, 0)
BUREEQUITY (0, 3) KINNEVIKINDB (0, 3) SECURITASB (2, 1)
CARDO (0, 2) KLIPPANS (3, 5) SEMCON (2, 1)
CASTELLUM (1, 0) KNOWIT (2, 3) SHBA (0, 1)
CLOETTAFAZERB (2, 1) LATOURINVESTMENTB (2, 1) SIGMAB (0, 0)
COLU B (4, 1) LEDSTIERNANB (0, 3) SINTERCASTA (0, 0)
CONCORDIAMARITIMEB (3, 4) LINDEX (0, 0) SKANDIA (3, 3)
CONNOVA (1, 4) LJUNGBERGGRUPPEN (4, 0) SKANSKAB (0, 4)
CONSILIUMB (0, 1) LUNDBERGSB (1, 0) SKFB (1, 4)
CUSTOSA (2, 0) MALD B (2, 2) SOFTRONICB (4, 3)
DILI (1, 2) MANDAMUS (2, 1) SSABA (1, 2)
DORO (1, 4) MAXIMPHARMSSEE (1, 3) STORAENSORSEE (0, 1)
DROTTB (1, 0) MEDITEAMDENTAL (2, 4) STRALFORSB (1, 0)
ELANDERSB (3, 0) MEDIVIRB (4, 5) SVEDBERGSB (0, 0)
ELECTROLUXB (0, 3) MODUL1DATA (6, 3) SVENSKAORIENTB (0, 2)
ELEKTAB (6, 6) MSCKONSULTB (0, 0) SVOLDERB (0, 0)
ELEKTRONIKGRUPPENB (4, 2) MUNKSJO (0, 1) SWEDISHMATCH (1, 1)
ENAT (0, 1) MUNTERS (2, 0) SWITCHCORE (0, 4)
ENEADATA (4, 2) NCCB (1, 2) SYD C (1, 2)
ENTR A (3, 2) NEFABB (4, 1) TELE2B (2, 6)
ERICSSONB (0, 3) NETINSIGHTB (0, 2) TELELOGIC (0, 4)
ESSELTEA (1, 4) NEWWAVEGROUPB (1, 0) TELIA (0, 1)
EUROPOLITAN (4, 3) NHNORDISKAB (1, 5) TELIGENT (1, 3)
FAGERHULT (2, 1) NOBELBIOCARE (7, 1) TICKETTRAVEL (2, 0)
FGLD (0, 3) NOKIASDBSEE (2, 5) TORNETFASTIGHETSB (0, 0)
FINNVEDENB (1, 3) NOLATOB (2, 4) TRELLEBORGB (3, 5)
FNSPKA (0, 2) NORDEA (0, 1) TURNITB (2, 5)
FORC B (0, 2) NORSKHYDROSDB (2, 0) TV4A (3, 2)
FRAMTIDSFABRIKEN (1, 3) OEMINTERNATIONALB (3, 0) UTFORS (0, 2)
FRONTECB (2, 1) OM (2, 4) VOLVOB (0, 4)
GAMBROA (1, 3) ORESUND (2, 1) VOSTOKNAFTASDB (2, 2)
GETINGEINDUSTRIER (2, 2) ORTIVUSB (3, 0) WALLENSTAMB (1, 3)
GEVEKOB (3, 0) OXIGENESEE (1, 2) WIHLBORGSB (1, 1)
WMDATAB (1, 4)
A APPENDICES 50
A.6 Some Basic Financial and Statistical Concepts

Adaptive Market Efficiency is a somewhat weaker form of market efficiency that allows
for the appearance of profit opportunities in historical data, but requires these profit
opportunities to dissipate when they become apparent.
Alpha According to the traditional capital asset pricing model, the expected rate of return
for asset i can, in equilibrium, be expressed as:
E[ri ] = rf + βi (E[rm ] − rf ),
where ri is the rate of return on asset i, βi is defined as below, rf is the risk-free rate of
return, and rm is the return on the market portfolio. Since rf is assumed to be risk-free,
it has zero variance, and can hence be handled as a constant, i.e. E[rf ] = rf . If the
expected rate of return exceeds this equilibrium value, then the asset is underpriced,
and we would expect it to outperform the market, adjusted for risk. We cannot directly
observe expectations, but we can infer them statistically from averages over historical
data. If we regress ri − rf against rm − rf , we assume that
ri − rf = αi + βi (rm − rf ) + εi , (18)
where εi ∼ N (0, s2 ) is the error term, and s2 is the sample variance. Taking the
expectation, we get
αi = E[ri ] − rf − βi (E[rm ] − rf ),
so we can estimate βi as the slope coefficient in the regression, and we can interpret
our estimate of the intercept αi as an estimate of the amount by which the ith asset
has outperformed the market (of course, if αi < 0, the security has underperformed in
relation to the market).
Arbitrage Informally, arbitrage means that an agent can act in such a way that she is
guaranteed a profit without exposure to risk, or with insignificant exposure. More
formally we can define price as the vector {St ; t = 0, 1}, if we assume that the market
only exists at time t = 0 and at t = 1. A portfolio is in this context defined as
h = (h1 , . . . , hN )T . We can define the value process V (h) as
Vt (h) = hT St , t = 0, 1
The portfolio h is an arbitrage strategy if it satisfies
V0 (h) ≤ 0, P (V1 (h) > 0) = 1
In the context of this thesis, an arbitrage opportunity emerges when a security for some
reason is mispriced. That is for example the case with the double listings of the Royal
Dutch/Shell stocks (they are listed both as Royal Dutch and as Shell, see e.g. Shleifer
(2000) for a discussion).
Autocorrelation Given returns R1 , R2 , . . . , RN , the lag k autocorrelation function is defined
as PN
(Ri − R̄)(Ri+k − R̄)
rk = 1 PN
2
1 (Ri − R̄)
As one of our performance measures, we use k = 1, 7, 30 lag(s). In words, autocorrela-

tion is the correlation between lagged realizations of a random variable.
A APPENDICES 51
Behavioral finance This attempt at explaining the movements of stock prices says that
”investors can make systematic errors in forecasting cash flows or in setting the discount
rate, and [that] these errors can push stock prices away from fundamental value for
extended periods of time.” This description is taken from Barberis (June 23 1997).
Beta β is a measure of the systematic risk in a security. The technical definition is
σim
βi = 2 ,
σm
where βi is the risk of security i, σim is the covariance between the security and the
2
market, and σm is the variance of the market.
CAR Abnormal return is the return on an asset in relation to some index. Cumulative
abnormal return is of course the cumulation of abnormal returns of some period of
time. In this thesis, we also use the acronym CAR as a reference for a simplistic
methodology (see below) where a market neutral portfolio is created and evaluated.
Correlation The correlation coefficient ρ indicates the extent to which pairs of realizations
of two stochastic variables lie on a straight line. For perfect linearity, ρ = ±1. If there
is no linear trend at all – for example, if there is a random scatter of points – the
value of ρ is close to zero. Points distributed evenly around a circle would also give a
correlation of near zero, because there would be no overall linear trend. The correlation
coefficient between the random variables X and Y is defined as
σXY
ρ= ,
σX σY
where σXY is the covariance between X and Y , and σX is the standard deviation of
X.
Data mining The data mining method is intended to work on data without starting from a
particular hypothesis or even a particular question. Essentially, it reverses the scientific
method, starting from data and moving towards hypotheses instead of following the
traditional order of the scientific method:
1. Define the problem
2. Generate hypotheses/models
3. Collect relevant data/conduct experiments to generate data
4. Test models against the data
5. Use the results to generate new hypotheses
The problems with this approach should be obvious to readers familiar with regression
analysis.
Data-snooping is finding seemingly significant but in fact spurious patterns in data. Ant-
onym for “our model”.
Derivative instruments Contracts such as options and futures whose price is derived from
the price of an underlying financial asset.
EMH (Efficient Market Hypothesis) This hypothesis says that the ”price (of e.g. a
stock) reflects fundamental value, defined as the best possible forecast, given available
information, of a security’s future cash flows, discounted at a rate that is appropriate
for the risk of those cash flows”.
Fat Tails If one uses the normal distribution to model stock returns, one will get too low
probabilities for extreme events, i.e. the mass in the tails are too low. One way to get
around this is to use e.g. hyperbolic distributions, which will render extreme events a
larger likelihood.
A APPENDICES 52
Fundamental data Data that is objective in some sense, e.g. accounting data.
Glamour stock A glamour stock is characterized by a high P/E-ratio, high trading volume
and long term positive earnings surprises. See Swaminathan and Lee (2000).
Growth stocks Companies that have a strong predicted growth in earnings, reflected in
e.g. a high P/E-ratio. The opposite of growth stock is value stock.
Hedge Hedging is a way of reducing some of the risk involved in holding an investment. A
hedge is a way of insuring an investment against risk.
Kurtosis Kurtosis is a measure of whether the data are peaked or flat relative to a normal
distribution. That is, data sets with high kurtosis tend to have a distinct peak near the
mean, decline rather rapidly, and have heavy tails. Data sets with low kurtosis tend to
have a flat top near the mean rather than a sharp peak. A uniform distribution would
be the extreme case. For the data Y1 , Y2 , . . . , YN , the formula for kurtosis is
PN
− Ȳ )4
i=1 (Yi
kurtosis = ,
(N − 1)s4
where Ȳ is the mean, s is the standard deviation and N is the number of data points.
Leverage The use of debt financing to increase the return of an investment.
Long position To be long in a security means to actually own it.
Market neutral A market neutral strategy is neutral to market movements, i.e. its β is
practically zero.
Markov Process A stochastic process x(t) is a Markov process if
P (x(tn ) ≤ xn |x(t)∀t ≤ tn−1 ) = P (x(tn ) ≤ xn |x(tn−1 )) ∀n, t1 < t2 < · · · < tn
In words, ”a Markov process models a situation in which where one is, is all one needs
[to] predict the future – how one got there provides no further information” (Bingham
and Kiesel, 1998).
Martingale A sequence of random variables X0 , X1 , . . . with finite means such that the
conditional expectation of Xn+1 given X0 , X1 , . . . , Xn is equal to Xn :
E[Xn+1 |X0 , X1 , . . . , Xn ] = Xn
Momentum Here defined as persistence in returns of stocks over approximately three

months to one year.
Neglected stock The opposite of ‘glamour stock.
P/E ratio The ratio of price to earnings for a stock.
Proxy Stand-in or substitute. Proxy data is used in statistics when it is not possible to
directly measure the object or phenomenon under observation.
Random Walk To get a mental picture of this concept, it is helpful to think about a
moving particle. In one dimension (the space of interest in this thesis) the particle
moves right with probability p and left with probability q = 1 − p. The jumps are made
independently of each other. A random walk is a martingale.
Risk Risk is subjective. In financial models, risk is often (e.g. in the mean-variance ap-
proach, CAPM etc) modelled with a quadratic utility function of the form U (x) =
A APPENDICES 53
x − bx2 , where U is the utility, x is a random wealth variable, and b is a positive para-
meter. But this is just one specification of the utility model, and there is an unlimited
number of different specifications. The quadratic utility function models a risk-averse
behavior, but there are also investors who are risk-lovers or risk-neutral, and their
utility functions are of course of another form.
Risk Averse Investor Risk aversion implies that when facing choices with comparable
returns, investors tend chose the less-risky alternative.
Risk Neutral Investor A risk neutral investor, choosing between different lotteries, picks
the one that gives the highest expected monetary income.
Semi-deductive approach To simultaneously use both empirical data and theories to
reach a conclusion about reality.
Sharpe ratio This ratio is defined as
rX − rf
S(X) = ,
σX
where X is some investment, rX is the average annual return on this investment, rf is
the best available return of a ”risk-free” security and σX is the standard deviation of
rX .
Short position Taking a short position means that one sells securities which one does not
own, and buys them back later, hopefully at a lower price than when one sold them.
Simplistic Methodology refers to the use of the classic CAR methodology. Stocks are
ranked by price change the last X months, the best say 10-20% and worst 10-20% of
the stocks available form the long and short portfolio respectively. These are equally
weighted. Then these stocks are kept for Y months until the portfolio is rebalanced.
Abnormal return in this market neutral portfolio is measured against 0% (which it
should yield if markets where efficient, save transaction costs and interest rate received
on the security deposit for the short positions). It is sufficient to use a program such
as Excel to do these kinds of simulations.
Simplistic Strategy in this thesis refers to a strategy that emulates the simplistic meth-
odology while using our simulation platform. We rank stocks after the price change the
last 6 months, and hold them for 4 months until they are all replaced simultaneously. 10
stocks are always included in the long and short portfolios respectively. No transaction
costs or other trading rules are used. The simulations with these parameter settings
will yield the same kind of risk-return profile as those presented in the empirical section
using CAR, as evidenced by our longer simulations in the Appendix.
Skewness For the data Y1 , Y2 , . . . , YN , the formula for skewness is:
PN
(Yi − Ȳ )3
skewness = i=1 ,
(N − 1)s3
where Ȳ is the mean, s is the standard deviation and N is the number of data points.
The skewness for a normal distribution is zero, and any symmetric data should have a
skewness near zero. Negative values for the skewness indicate data that are skewed left
and positive values for the skewness indicate data that are skewed right. By skewed
left, we mean that the left tail is heavier than the right tail. Similarly, skewed right
means that the right tail is heavier than the left tail.
Trading horizon This is the end of the trading period, when the investment is assumed to
pay off.
A APPENDICES 54
t-statistic The t statistic is used to test hypotheses about an unknown population mean µ
in situations where the value of σ is unknown. Often used on e.g. monthly return or
CAR when using the simplistic methodology, where the null hypothesis is that returns
are zero. It is however afflicted with problems in our case since the return distribution
of a market neutral momentum strategy has fatter tails than the one assumed when
using the t-statistic. We believe that the performance measures used in this thesis give
a better picture.
Utility Function Quoting from Luenberger (1998): ”A utility function is a function U
defined on the real numbers (representing possible wealth levels) and giving a real value.
Once a utility function is defined, all alternative random wealth levels are ranked by
evaluating their expected utility values.” It is important to note that utility functions
do not produce cardinal values, so their only significance is in relation to other utility
scores. An important insight when reading this thesis is that a certain Sharpe value
can have different utility values, depending on the level of risk aversion.
Value stocks Companies which earnings are not assumed to grow fast in the future, reflec-
ted in e.g. a low P/E-ratio.
Zero-cost investment An investment that does not require capital to be paid upfront. A
margin (risk buffer) is however usually required. Going short in a stock is an example.
REFERENCES 55
References
Interview with Björn Germer, Head of SE-Banken Hedge Fund Equity, April 2002.
R. A. Adams. Calculus: A Complete Course. Addison-Wesley, Ontario, 3rd edition, 1995.
N. Barberis. Markets: the Price May Not Be Right. Financial Times, June 23 1997.
N. Barberis and A. Shleifer. A Model of Investor Sentiment. Working paper, NBER Working
Paper 5926, February 1997.
N. H. Bingham and R. Kiesel. Risk-Neutral Valuation. Springer, London, 1998.
W. F. M. De Bondt and R. Thaler. Does the Stock Market Overreact? Journal of Finance,
40(3):793–805, 1985.
W. Bondt D. Schiereck and M. Weber. Contrarian and Momentum Strategies in Germany.

Financial Analysts Journal, November/December, 1999.
K. Daniel and S. Titman. Market Efficiency in an Irrational World. Financial Analysts

Journal, November/December, 1999.
J. Bradford de Long, A. Shleifer, L. H. Summers, and R. J. Waldmann. Positive Feedback

Investment Strategies and Destabilizing Rational Speculation. Journal of Finance, 45(2):
379–395, 1990.
B. Djehiche. Stochastic Calculus: An Introduction with Applications. Royal Institute of

Technology, Stockholm, 2000.
E. Fama. Efficient Capital Markets: A Review of Theory and Empirical Work. Journal of
Finance, 48:383–417, March 1970.
E. Fama. Market Efficiency, Long-term Returns, and Behavioral Finance. Journal of Fin-
ancial Economics, 49:283–306, 1998.
E. Fama and K. French. Common Risk Factors in the Returns on Stocks and Bonds. Journal
of Financial Economics, 33:3–56, 1993.
E. Fama and K. French. Multifactor Explanations of Asset Pricing Anomalies. Journal of

Finance, 51:55–84, 1996.
M. Grinblatt and B. Han. The Disposition Effect and Momentum. Working Paper 8734,
NBER, January 2002.
A. Gut. An Intermediate Course in Probability. Springer, New York, 1995.
H. Hong and J. Stein. A Unified Theory of Underreaction, Momentum Trading and Overre-
action in Asset Markets. Journal of Finance, 54:2143–2184, 1999.
A. Shleifer J. Lakonishok and R. Vishny. Contrarian Investment, Extrapolation, and Risk.

Journal of Finance, 49:1541–1578, 1994.
N. Jegadeesh and S. Titman. Returns to Buying Winners and Selling Losers: Implications
for Stock Market Efficiency. Journal of Finance, 48:65–93, March 1993.
REFERENCES 56
N. Jegadeesh and S. Titman. Profitability of Momentum Strategies: An Evaluation of

Alternative Explanations. Journal of Finance, 56:699–720, September 2001.
J.D. Jobson and B. Korkie. Performance Hypothesis Testing with the Sharpe and Treynor
Measures. Journal of Finance, 36:888–908, 1981.
D. Kahneman and A. Tversky. Prospect Theory: An Analysis of Decision under Risk.

Econometrica, 47(2):263–292, March 1979.
M. Könberg and M. Lindberg. Strategic Investing in Hedge Funds. Master thesis, Linköpings
Universitet, 2000.
N. Jegadeesh L.K.C. Chan and J. Lakonishok. The Profitability of Momentum Strategies.

Financial Analysts Journal, November/December, 1999.
D. G. Luenberger. Investment Science. Oxford University Press, New York, 1998.
R. Fuller M. Finn and J. Kling. It’s mostly on the short side. Financial Analysts Journal,
November/December, 1999.
B. Mandelbrot. Forecasts of Future Prices, Unbiased Markets and "Martingale" Models.

Journal of Business, 39(1):242–255, 1966.
P. Newbold. Statistics for Business and Economics. Prentice Hall, Englewood Cliffs, 1995.
E. S. O’Neal. Industry Momentum and Sector Mutual Funds, Financial Analysts Journal.
Financial Analysts Journal, July/August, 2000.
J.D. Piotroski. Value Investing: The Use of Historical Financial Statement Information to
Separate Winners from Losers. Journal of Accounting Research, 38, 2000.
A. Shleifer R. La Porta, J. Lakonishok and R. Vishny. Good News for Value Stocks: Further
Evidence on Market Efficiency. Journal of Finance, 52:859–874, 1997.
K. G. Rouwenhorst. International Momentum Strategies. Journal of Finance, 53:267–284,

1998.
L. Råde. Mathematics Handbook for Science and Engineering. Springer, Berlin, 1999.
R. J. Shiller. Human behavior and the Efficiency of the Financial System. Working Paper
6375, NBER, January 1998.
R. J. Shiller. Irrational Exuberance. Princeton University Press, New Jersey, 2000.
A. Shleifer. Inefficient Markets, An Introduction to Behavioral Finance. Oxford University

Press, New York, 2000.
B. Swaminathan and C.M.C. Lee. Price Momentum and Trading Volume. Working paper,
Cornell University, 1999.
B. Swaminathan and C.M.C. Lee. Do Stock Prices Overreact to Earnings News? Working
paper, Johnson Graduate School of Management, January 2000.
P. Säfvenblad. Creating Returns and Controlling Risks in Managed Funds. Lecture notes,
Stockholm School of Economics, 2001.
REFERENCES 57
J. Söderström. The Swedish Momentum Effect. Master thesis, Stockholm School of Eco-
nomics, 2000.

Larsson Larsson Aberg Statistical Arbitrage Final Version 2002

Hochgeladen von

Dokumentinformationen

Originalbeschreibung:

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Larsson Larsson Aberg Statistical Arbitrage Final Version 2002

Hochgeladen von

Copyright:

Verfügbare Formate

A Market Neutral Statistical Arbitrage Trading Model

7 Development of Our Default Model

13th March 2003

3 An Overview of Our Model and Its Context 19

5 Methodology – Signal Generation and Risk Control 25

6 Results - Portfolio Simulations 29

1.4 Outline and Reader’s Guide

2.1 Theoretical Justification for Momentum Strategies

2.1.1 Efficient Markets

E(p̃j,t+1 |Φt ) = [1 + E(r̃j,t+1 |Φt )]pj,t , (1)

2.1.2 Data mining and Spurious Data

2.2 A Model of Over- and Underreaction

where Nt is the earnings in period t and

which is a geometric series, and can be rewritten as

processes (see Appendix A.6 for an explanation of Markov processes).

((1 − λ1 )qt + λ2 (1 − qt ))P (yt+1 |st+1 = 1, yt )

"True" Value of Stock

P (−4.15 · 10−3 < µ < −4.09 · 10−3 ) = 0.99,

µ ∈ (2.62 · 10−4 , 3.06 · 10−4 )

2.2.1 Conclusions and Discussion

2.3 Theoretical Motivation for the Stop-Loss Rule

accumulated gain up to the nth toss, S0 = 0. Then Sn evolves according to

2.4 Empirical Evidence Using CAR

2.4.1 The Momentum Effect

Table 2. A brief summary of empirical research on the momentum effect.

Paper Data Sample Results Other Conclusions

O’Neal (2000) S&P (1989-1999) Much of the observed momentum in individual

One-month rank period

2.4.2 The Role of Volume, Fundamental and Sentiment Variables

3 An Overview of Our Model and Its Context

Table 5. Hedge fund history at a glance. Reproduced from Säfvenblad (2001).

Asset allocation within hedge fund universe in 2000:

Hedge Fund Strategies

Table 7. Definition of Statistical Arbitrage.

An approach is defined to be Statistical Ar-

• positions are entered into on the basis of

3.2 Our Model – an Optimization Problem

Table 9. Our strategy.

Aspect of Choice Our Choice Justification

odology – Signal Generation and Risk Control).

3.3 Technical Specification of the Computer Program

Figure 5. Conceptual view of the computer system.

4.1 Description of the Data

Table 11. The stocks used in the model.

used please see Table 23 in Appendix A.5.

5 Methodology – Signal Generation and Risk Control

5.1 Signal Generation

Table 12. Our momentum-proxies.

Characteristics of strong momentum Our Proxy

More susceptible to sentiment changes, i.e. High M/B

Positive/negative sentiment in trading Large changes in price during increase in

Fundamental strength/weakness Posit- Large changes in analyst expectations of

5.2 Risk Control

5.3 Assumptions in Our Simulations

5.3.1 Basic Assumptions of Our Trading Model

6 Results - Portfolio Simulations

7 Development of Our Default Model

Table 13. Evaluation variables.

Evaluation Variable Mathematical Definition Comment on variable

Daily Return day t Total Value

Downside Risk Pt=T 2 This measure gives a value to the aggreg-

Beta (AFGX) See Appendix Comovement with market. A low β provides

Accumulated transaction This measure quantifies total transaction